An underground traffic engineering earthquake damage dynamic prediction method and system

Abstract

The application relates to an underground traffic engineering earthquake damage dynamic prediction method and system, solves the core problems of poor adaptability of a static model, multi-agent interaction conflict and insufficient real-time performance of bidirectional coupling in underground traffic engineering earthquake damage dynamic prediction, and the method comprises the following steps: constructing an earthquake disaster panoramic knowledge graph containing a semantic relationship ontology model of a disaster-bearing body, a disaster-causing factor and a disaster-forming environment; establishing a multi-agent simulation model containing structure, facility, environment and personnel agents; building a bidirectional coupling mechanism (rules are mapped into parameter constraints, real-time data collection is fed back to update the knowledge graph); inputting seismic motion parameters, performing semantic reasoning on the knowledge graph and parallel simulation on the multi-agent, dynamically predicting service performance loss, and outputting structure damage and functional failure quantitative evaluation results. The application has the following effects: bidirectional dynamic coupling of the knowledge graph and the multi-agent is realized, interaction conflicts are eliminated, damage is accurately predicted in real time under dynamic evolution of a scene, and low-cost emergency decision-making is supported.

Description

Technical Field

[0001] This invention relates to the field of earthquake resistance and disaster prevention in underground engineering, and in particular to a method and system for dynamic prediction of earthquake damage in underground transportation engineering. Background Technology

[0002] Underground transportation engineering is a critical infrastructure for urban lifelines, and its seismic safety is directly related to public safety and economic operation. Under earthquake action, these projects face multiple threats, including accumulated structural damage, triggering of secondary disaster chains, and obstruction of personnel evacuation. Their service performance exhibits dynamic degradation characteristics, necessitating dynamic prediction methods that can reflect the evolution of disasters.

[0003] In related technologies, earthquake damage prediction mainly adopts deterministic assessment methods based on numerical simulations such as finite element method or empirical statistics. By pre-setting seismic motion input and failure criteria, the structural response and damage level are calculated offline.

[0004] However, such methods belong to the open-loop static analysis paradigm, which cannot capture the suddenness and chain evolution of earthquake disaster chains, nor can they reflect the real-time dynamic coupling between surrounding rock, structure, facilities and personnel. This results in a serious disconnect between the prediction results and the continuously changing service status during the disaster, making it difficult to support the demand for dynamic prediction accuracy for second-level decision-making during a disaster. Summary of the Invention

[0005] To achieve bidirectional dynamic coupling between knowledge graphs and multiple agents, resolve interaction conflicts, and accurately predict damage in real-time under dynamic evolution scenarios, thereby supporting low-cost emergency decision-making, this application provides a method and system for dynamic prediction of earthquake damage in underground transportation engineering.

[0006] Firstly, this application provides a method for dynamic prediction of seismic damage in underground transportation engineering projects, employing the following technical solution:

[0007] A method for dynamic prediction of seismic damage in underground transportation engineering projects includes:

[0008] Based on data reflecting the characteristics of earthquake disasters in underground transportation projects, a panoramic knowledge graph of earthquake disasters in underground transportation projects is constructed. This knowledge graph includes a schema layer and a data layer. The schema layer adopts an ontology to describe the disaster-bearing body, disaster-causing factors, disaster-generating environment and the semantic relationships between them in the earthquake disaster environment. The data layer contains instance data corresponding to the ontology.

[0009] Based on the constructed knowledge graph, a multi-agent simulation model for underground transportation engineering is established. The model includes a structural agent that simulates the mechanical response of the structure, a facility agent that simulates the functional state of the facility, an environmental agent that simulates the propagation of seismic motion, and a personnel agent that simulates the emergency behavior of personnel. Each agent interacts according to the rules provided by the knowledge graph.

[0010] A two-way coupling mechanism is established between the knowledge graph and the multi-agent simulation model. The semantic rules in the knowledge graph are mapped to the initialization parameters and behavioral constraints of the simulation model. At the same time, the state change data during the simulation process is collected in real time and fed back to the knowledge graph for dynamic updates.

[0011] Based on a knowledge graph and multi-agent simulation model with a two-way coupling mechanism, seismic motion parameters are input, and the service performance loss state of underground transportation engineering is dynamically predicted through semantic reasoning of the knowledge graph and parallel simulation of the multi-agent model. The quantitative assessment results of structural damage and functional failure are output.

[0012] Secondly, this application provides a dynamic prediction system for earthquake damage in underground transportation engineering projects, employing the following technical solution:

[0013] A dynamic earthquake damage prediction system for underground transportation engineering includes a memory, a processor, and a program stored in the memory and executable on the processor. When the program is loaded and executed by the processor, it can implement the dynamic earthquake damage prediction method for underground transportation engineering as described in the first aspect. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating a dynamic prediction method for earthquake damage in underground transportation engineering according to an embodiment of this application. Detailed Implementation

[0015] The present application will be further described in detail below with reference to the accompanying drawings.

[0016] Reference Figure 1 This application discloses a method for dynamic prediction of seismic damage in underground transportation engineering, comprising:

[0017] Step S1: Based on data reflecting the characteristics of earthquake disasters in underground transportation engineering, construct a panoramic knowledge graph of earthquake disasters in underground transportation engineering. The knowledge graph includes a schema layer and a data layer. The schema layer uses an ontology to describe the disaster-bearing body, disaster-causing factors, disaster-prone environment, personnel and their semantic relationships in the earthquake disaster environment. The data layer contains instance data corresponding to the ontology. The instance data includes instance attributes and semantic relationships between instances. The semantic relationships have an initial association strength defined according to the ontology.

[0018] Among them, the panoramic knowledge graph of earthquake disasters refers to a data organization method that uniformly represents the earthquake disaster elements and their relationships in the form of a graph structure. In this application, it specifically refers to a semantic network covering four core elements: disaster-bearing body, disaster-causing factor, disaster-prone environment and personnel, which is used to support subsequent semantic reasoning and multi-agent simulation.

[0019] The schema layer refers to the conceptual framework layer of the knowledge graph. In this application, it specifically refers to the class hierarchy and attribute constraint set constructed based on ontology methods, which is used to define the formal semantics of concepts in the earthquake disaster domain.

[0020] The data layer refers to the instance layer of the knowledge graph, specifically the collection of entity instances and their attribute values ​​corresponding to various ontology concepts in the schema layer. An ontology refers to a shared, conceptualized, and explicitly formalized specification, specifically a domain conceptual model of disaster-bearing bodies, disaster-causing factors, disaster-prone environments, personnel, and their semantic relationships in an earthquake disaster environment, formally expressed using Web Ontology Language (OWL). Semantic relationships refer to the logical connections between entities defined in the ontology, specifically the disaster chain logical relationships such as disaster-causing factors acting on disaster-bearing bodies, the disaster-prone environment modulating the intensity of disaster-causing factors, and personnel trapped in disaster-bearing bodies. Initial association strength refers to the weight values ​​assigned to semantic relationships in the early stages of knowledge graph construction based on historical earthquake damage statistics or expert experience; specifically, it refers to a dimensionless scalar with a value range of [0,1], used to quantify the tightness of the connections between disaster elements.

[0021] The necessary procedures are as follows:

[0022] Step S1-1 involves collecting multi-source heterogeneous data reflecting the seismic hazard characteristics of underground transportation engineering projects. This data includes: structural as-built drawings and material strength test reports (static properties of the disaster-bearing body), historical ground motion records and site soil dynamic parameters (hazard-causing factors and disaster-prone environment), and personnel distribution and evacuation plans within the underground transportation engineering project (personnel attributes). For example, for a subway tunnel section, data such as the lining concrete strength grade C35, surrounding rock level III, peak historical ground acceleration of 0.20g, and daily peak-hour personnel density of 1.2 people / m² are collected.

[0023] Steps S1-2: Construct a model layer ontology based on the collected data. An ontology editing tool is used to establish a five-tuple ontology structure O=(C,P,R,A,H), where C is the concept set {disaster-bearing body, disaster-causing factor, disaster-inducing environment, personnel}, P is the attribute set, R is the relation set {inducing, acting on, modulating, trapped, evacuated to}, A is the axiom set, and H is the hierarchical structure. For example, the concept "tunnel lining structure" is defined as a subclass of the disaster-bearing body, with attributes including lining thickness (unit: m), concrete compressive strength (unit: MPa), and crack width (unit: mm); the attribute "damage level" is defined as an ordered categorical variable, with a value range of {slight, moderate, severe, collapse}, corresponding to damage index intervals of [0,0.3), [0.3,0.6), [0.6,0.9), and [0.9,1.0], respectively.

[0024] Steps S1-3 map the collected data into data layer instances. Based on the ontology definition of the model layer, instantiate specific engineering objects: create an instance of "Left Line of Tunnel in XX Section", classify it under the concept of "Tunnel Lining Structure", and assign attribute values ​​{lining thickness: 0.35m, concrete compressive strength: 35MPa, crack width: 0mm}; create an instance of "Historical Earthquake Event M", classify it under the concept of "Disaster Causing Factor", and assign attribute values ​​{peak acceleration: 0.20g, duration: 12s, spectral characteristics: soft soil site}.

[0025] Steps S1-4 establish semantic relationships between instances and assign initial association strengths. Based on disaster chain logic and historical earthquake damage statistics, directed edges are established between instances and assigned initial weights. For example, a semantic relationship is established between "historical earthquake event M" → "acts on" → "left line of tunnel in section XX", with an initial association strength of 0.75, representing the typical probability of damage to the tunnel structure by the historical earthquake; a semantic relationship is established between "left line of tunnel in section XX" → "trapped by" → "peak period personnel distribution", with an initial association strength of 0.90, representing the high correlation between structural damage and personnel being trapped.

[0026] Step S2: Based on the knowledge graph, establish a multi-agent simulation model for underground transportation engineering. The multi-agent simulation model includes a structural agent simulating structural mechanical response, a facility agent simulating facility functional state, an environmental agent simulating seismic propagation, and a personnel agent simulating emergency behavior. Each agent interacts according to the rules provided by the knowledge graph.

[0027] Among them, the multi-agent simulation model refers to a distributed simulation system in which multiple computational entities with autonomous decision-making capabilities interact and generate emergent behaviors in a shared environment. In this application, it specifically refers to a computational framework for simulating the seismic response and emergency evolution process of underground transportation engineering. The structural intelligent agent refers to a computational entity that simulates the structural mechanical response of underground transportation engineering. In this application, it specifically refers to an intelligent agent unit with damage accumulation and stiffness degradation calculation capabilities at the granularity of finite element nodes or structural components, whose state space includes displacement, stress, crack width, and structural damage index. The facility intelligent agent refers to a computational entity that simulates the functional state of underground transportation engineering facilities. In this application, it specifically refers to an intelligent agent unit with functional degradation and resource scheduling calculation capabilities at the granularity of ventilation, lighting, drainage, and communication subsystems, whose state space includes facility operating status, resource reserves, and functional availability. An environmental intelligent agent refers to a computational entity that simulates seismic motion propagation and site response. In this application, it specifically refers to an intelligent agent unit with the granularity of seismic wave input and soil-structure interaction, possessing the ability to generate and propagate seismic fields. Its state space includes seismic acceleration time history, soil shear wave velocity, and seismic intensity index. A personnel intelligent agent refers to a computational entity that simulates the emergency behavior of personnel inside underground transportation projects. In this application, it specifically refers to an intelligent agent unit with the granularity of individuals or groups, possessing the ability to select evacuation routes, perform disaster avoidance behaviors, and calculate panic propagation. Its state space includes personnel location coordinates, movement speed, panic index, and evacuation status.

[0028] The necessary procedures are as follows:

[0029] Step S2-1: Extract initialization parameters for the intelligent agent from the knowledge graph. Based on the disaster-bearing body instance attributes in the knowledge graph data layer, extract structural geometric parameters and material mechanical parameters as initialization inputs for the structural intelligent agent; based on the facility instance attributes, extract the facility's rated power, redundancy, and resource capacity as initialization inputs for the facility intelligent agent; based on the disaster-causing factor and disaster-prone environment instance attributes, extract seismic motion parameters and site soil parameters as initialization inputs for the environmental intelligent agent; based on the personnel instance attributes, extract the number of personnel, distribution density, and evacuation plan as initialization inputs for the personnel intelligent agent. For example, extract the lining thickness of 0.35m and elastic modulus of 30GPa for the left tunnel of section XX from the knowledge graph and assign it to the structural intelligent agent as the basis for initial stiffness calculation.

[0030] Step S2-2: Construct the internal state transition model for each agent. The structural agent uses an elastoplastic damage mechanics model to describe the state transition, with the state equation being: ,in This is the stress tensor (unit: MPa). The structural damage index (value range [0,1]) Elastic modulus (unit: GPa). For the strain tensor; the facility agent uses a Markov state transition model to describe the functional state evolution, with a state space of {normal, degraded, failed}, and the transition probabilities are calibrated based on historical statistics of facility failure rates; the environmental agent uses a numerical solution method of the wave equation to describe seismic motion propagation, with the governing equation being... ,in This refers to the soil density of the site (unit: kg / m³). This represents the relative displacement of the soil (unit: m). Shear modulus (unit: Pa). This represents the bedrock ground motion displacement (unit: m). Time (unit: seconds) Let be the depth coordinate (m). The above equation is the standard form of the one-dimensional wave equation for the site soil-bedrock system. (Left side) The soil inertial force density (N / m³); the first term on the right. The soil shear recovery force density (N / m³); the second item on the right. The equivalent volumetric force density (N / m³) of the input seismic motion to the bedrock characterizes the driving effect of bedrock movement on the overlying soil. All three quantities are dimensionless (N / m³), and the equations are dimensionally homogeneous. Technicians can use the finite difference method or the finite element method to discretize and solve the time-space domain, with boundary conditions taken at the bedrock location. Free surface at the Earth's surface The human agent uses a social force model to describe its movement behavior, and the equation of motion is: ,in Personnel weight (unit: kg) Personnel speed (unit: m / s) Driving force (unit: N). Repulsive force between people (unit: N). Repulsive force of environmental obstacles (unit: N).

[0031] Steps S2-3 establish communication and interaction interfaces between intelligent agents. Define message formats and communication protocols for four types of intelligent agents: The environment agent broadcasts seismic acceleration time histories to the structure agent (message type: environment-structure, data format: time-acceleration sequence, unit: m / s). 2The structural agent sends structural damage status to the facility agent (message type: structure-facility, data format: damage index-location coordinates); the facility agent broadcasts facility functional status and resource reserves to the personnel agent (message type: facility-personnel, data format: functional status-resource quantity); the personnel agent reports personnel location and load to the structural agent (message type: personnel-structure, data format: location coordinates-equivalent load). For the specific execution process of interactions between agents based on the rules provided by the knowledge graph, please refer to steps S21 to S25; for the specific execution process of parallel interactions between agents through hierarchical multi-agent reinforcement learning based on a hierarchical behavior policy library, please refer to steps S241 to S244. Further details are omitted here.

[0032] Step S3: Establish a bidirectional coupling mechanism between the knowledge graph and the multi-agent simulation model. Transform the semantic rules in the knowledge graph that reflect the evolution logic of the disaster chain into the initialization parameters and behavioral constraints of the multi-agent simulation model. The initialization parameters include structural mechanical parameters, facility status parameters, and personnel distribution parameters. At the same time, collect state change data in real time during the simulation process through a preset event-driven architecture, feed it back to the knowledge graph, and update the association strength of instance attributes and semantic relationships.

[0033] The bidirectional coupling mechanism refers to the technical architecture that forms a data closed loop between the knowledge graph and the multi-agent simulation model. In this application, it specifically refers to a bidirectional data path in which semantic rules drive simulation initialization and behavioral constraints in the forward direction, and the simulation state dynamically updates the attributes and association strength of knowledge graph instances in the reverse direction. Initialization parameters refer to the initial state quantities required when the multi-agent simulation model starts. In this application, they specifically refer to the structural mechanics parameters, facility state parameters, and personnel distribution parameters obtained by mapping semantic features from the knowledge graph. Behavioral constraints refer to the set of rules that restrict the feasible domain of agent actions. In this application, they specifically refer to the physical constraints, resource constraints, and safety constraints obtained by decoding semantic relationships from the knowledge graph. The event-driven architecture refers to a technical framework for data acquisition and processing based on state change events. In this application, it specifically refers to an asynchronous computing architecture that uses structural damage exceeding limits, personnel path deviation, and insufficient facility resources as triggering conditions to drive dynamic updates of the knowledge graph. State change data refers to the time-varying state quantities generated by each agent during the multi-agent simulation process. In this application, it specifically refers to dynamic indicators such as structural damage index, facility resource reserves, and personnel position coordinates. Incremental learning algorithms refer to learning mechanisms that continuously update model parameters based on newly arriving data without forgetting existing knowledge. In this application, it specifically refers to online gradient descent or Bayesian update methods used to dynamically update the attribute values ​​and association strengths of knowledge graphs.

[0034] The necessary procedures are as follows:

[0035] Step S3-1: Perform forward coupling data transformation. A preset graph neural network algorithm extracts multi-dimensional features of semantic rules from the knowledge graph. These extracted features are then converted into initialization parameters and behavioral constraints for a multi-agent simulation model, achieving a forward mapping from knowledge graph semantic rules to the initial state and action boundaries of the simulation model.

[0036] Step S3-2: Perform reverse coupling data feedback. Based on a preset event-driven architecture, real-time data on state changes generated by each agent during the operation of the multi-agent simulation model is collected. The state change data is processed by an incremental learning algorithm controlled by a time window, dynamically updating the attribute values ​​of instances corresponding to the ontology in the knowledge graph, adjusting the correlation strength of semantic relationships between instances, and realizing dynamic feedback updates from simulation state to the knowledge graph.

[0037] Step S3-3: Complete the establishment and verification of bidirectional coupling. Connect the forward and reverse coupling paths to verify the integrity and consistency of bidirectional data transmission, and establish the bidirectional coupling state between the knowledge graph and the multi-agent simulation model. For the specific execution process of graph neural network feature extraction and parameter mapping, refer to step S31; for the specific execution process of event-driven architecture real-time acquisition and incremental learning dynamic updates, refer to steps S32 to S3F; for the specific execution process of update validity verification, refer to steps S33 and S3E1 to S3E4. Further details are omitted here.

[0038] Step S4: After the bidirectional coupling mechanism is established, the seismic motion parameters are input based on the knowledge graph and the multi-agent simulation model. Semantic reasoning and parallel simulation of the multi-agent simulation model are performed through the knowledge graph to dynamically predict the service performance loss state of underground transportation engineering and output the quantitative assessment results of structural damage and functional failure.

[0039] Among them, seismic motion parameters refer to the set of physical quantities describing the characteristics of seismic motion, specifically peak ground acceleration, significant duration, site characteristic period, and input acceleration time history in this application. Semantic reasoning refers to the process of logical deduction based on knowledge graph ontology and rules, specifically the computational process of forward chain deduction performed by the reasoning engine with the current simulation state as fact and production rules as premise. Parallel simulation refers to a distributed computing mode in which multiple agents synchronously execute decisions and actions within the same time step, specifically the simulation mechanism in this application where four types of agents—structure, facility, environment, and personnel—coordinate and advance according to a unified clock. Service performance loss state refers to an index state that comprehensively reflects the degree of earthquake damage to underground transportation engineering, specifically the dimensionless scalar obtained by weighted aggregation of structural damage, facility failure, and personnel casualty risk in this application. Structural damage refers to the degradation of the mechanical properties of underground transportation engineering structural components under seismic loading, specifically the degree of bearing capacity reduction quantified by the structural damage index in this application. Functional failure refers to the loss of service capacity of underground transportation engineering facility systems during earthquake disasters, specifically the degree of operational interruption characterized by a decrease in functional availability in this application.

[0040] The necessary procedures are as follows:

[0041] Step S4-1: Input the seismic motion parameters and activate the knowledge graph instance. Based on the ontology in the knowledge graph, analyze the semantic relationships between the seismic motion parameters and disaster-causing factors, disaster-prone environment, disaster-bearing bodies, and personnel, activate the instances corresponding to the semantic relationships, and extract the attribute data of the activated instances as the initial conditions for simulation.

[0042] Step S4-2: Generate a dynamic prediction rule set. Based on the activated instances, key semantic rules are extracted using a pre-defined graph neural network algorithm to generate a dynamic prediction rule set suitable for the current earthquake scenario, which is used to drive the interaction behavior of multiple agents.

[0043] Step S4-3: Start parallel simulation. Start the multi-agent simulation model for parallel simulation. The environmental agent synchronously inputs seismic parameters to drive the simulation timing. The structural agent, facility agent, environmental agent, and personnel agent interact according to the dynamic prediction rule set. The state of each agent is updated synchronously through a shared clock.

[0044] Step S4-4: Real-time updates of prediction rules during simulation. State change data generated by each agent during simulation is collected in real time. Based on this data, semantic reasoning is performed using a knowledge graph to update the dynamic prediction rule set, ensuring that the simulation state and prediction rules are updated synchronously to adapt to sudden changes in disaster evolution.

[0045] Steps S4-5 aggregate and evaluate the quantitative results. Based on the updated dynamic prediction rule set, aggregate the state change data output by each agent during the simulation, calculate the service performance loss index, and output the quantitative evaluation results of structural damage and functional failure. For the specific execution process of activating instances and generating the dynamic prediction rule set, refer to steps S41 to S42; for the specific execution process of parallel simulation driving, refer to step S43; for the specific execution process of real-time rule updating, refer to steps S44 and S441 to S445; for the specific execution process of performance loss aggregation evaluation, refer to step S45; for the specific execution process of mode layer and interaction topology co-evolution, refer to steps S44A to S44E. These will not be elaborated upon here.

[0046] The interaction between agents based on rules provided by the knowledge graph is a crucial link connecting static disaster knowledge representation with dynamic simulation behavior generation. Existing technologies often rely on fixed scripts for agent interaction, making it difficult to adapt to the spatiotemporal evolution of disaster chains. This application achieves a triple leap in agent interaction: from pre-set scripts to data-driven interaction, from single-layer flatness to multi-layer collaboration, and from independent decision-making to conflict coordination. This is accomplished by extracting semantic rules for earthquake disasters, calculating spatiotemporal correlation weights, constructing a hierarchical behavior policy library, executing hierarchical multi-agent reinforcement learning, and introducing a game theory negotiation mechanism. The following details each sub-step.

[0047] The agents interact according to the rules provided by the knowledge graph, including:

[0048] Step S21: Extract earthquake disaster semantic rules from the rules provided by the knowledge graph to obtain a rule set; earthquake disaster semantic rules reflect the disaster-causing factors, disaster-inducing environment, disaster-bearing bodies and people, and the disaster-causing relationships constitute the semantic basis of the disaster chain.

[0049] Among them, earthquake disaster semantic rules refer to disaster chain logic rules formally expressed in the form of knowledge graph ontology. In this application, they specifically refer to production rules or descriptive logic axioms that reflect the disaster-causing factors, disaster-inducing environment, disaster-bearing bodies, and personnel. The rule set refers to a subset of semantic rules extracted from the knowledge graph that are applicable to the current simulation scenario. In this application, it specifically refers to a disaster chain rule library that has been screened and used to drive multi-agent interaction.

[0050] The necessary procedures are as follows:

[0051] Step S21-1: Traverse all semantic relation definitions related to earthquake disasters in the knowledge graph pattern layer and extract the disaster chain logic expressed in the form of production rules. For example, extract rule R1: "If the peak ground acceleration of the disaster-causing factor is greater than 0.15g (approximately 1.47m / s²), and the site category of the disaster-prone environment is soft soil, then the damage level of the tunnel structure in the disaster-bearing body is determined to be moderate or above. This rule characterizes the typical damage mode of strong earthquakes to tunnel structures under soft soil site conditions; where 0.15g is calibrated according to the lower limit of the intensity of frequent earthquakes in soft soil sites, and technicians can adjust it according to the site safety assessment report." This rule characterizes the typical damage mode of strong earthquakes to tunnel structures under soft soil site conditions.

[0052] Step S21-2: Instantiate, match, and filter the rule set based on the parameters of the current simulation scenario. Eliminate rules irrelevant to the current engineering scenario and retain rules with a matching degree higher than a preset threshold (e.g., 0.6). For example, if the current scenario is a hard soil site (shear wave velocity > 500 m / s) and peak ground acceleration 0.20 g, then rule R1 is eliminated due to site condition mismatch; rule R2 (example) is retained: "If the peak ground acceleration of the disaster-causing factor is greater than 0.15 g (approximately 1.47 m / s²), and the site type of the disaster-prone environment is hard soil, then the damage level of the tunnel structure in the disaster-bearing body is determined to be minor or above. This rule characterizes the initial damage mode of a moderate-intensity earthquake to the tunnel structure under hard soil site conditions; where 0.15 g is calibrated according to the lower limit of the intensity of frequent earthquakes in hard soil sites, and technicians can adjust it according to the peak ground acceleration parameter in the site seismic safety assessment report."

[0053] Step S21-3: The filtered rule set is encoded into standardized rule objects and stored in the decision rule cache area of ​​the hierarchical behavior policy library for subsequent processing by the spatiotemporal graph attention network. The data structure of the rule object includes: rule identifier (string), set of preconditions (list of Boolean expressions), conclusion action (string), and confidence score (dimensionless, value range [0,1]).

[0054] Step S22: Calculate the spatiotemporal correlation weights of each rule in the rule set using a preset spatiotemporal graph attention network, obtain weighted rules based on the spatiotemporal correlation weights, and identify key disaster chains with spatiotemporal correlations based on the spatiotemporal correlation weights.

[0055] In this application, the spatiotemporal graph attention network refers to a deep learning model that integrates graph neural networks and attention mechanisms, simultaneously modeling spatial topological relationships and temporal dynamic evolution. Specifically, it refers to a network structure that uses disaster rules as nodes, spatiotemporal coupling relationships between rules as edges, and calculates rule importance through multi-head attention. The spatiotemporal correlation weight is a scalar value that quantifies the degree of correlation between two disaster rules in terms of spatial proximity and temporal sequence. Specifically, it refers to the dimensionless weight after normalization of the attention coefficients in the spatiotemporal graph attention network, with a value range of [0,1]. The critical disaster chain refers to a disaster evolution path formed by sequentially linking highly correlated disaster rules. Specifically, it refers to a sequence of rules with a spatiotemporal correlation weight exceeding a preset threshold (e.g., 0.7), used to identify the main channel of disaster propagation.

[0056] The necessary procedures are as follows:

[0057] Step S22-1: Construct a rule spatiotemporal graph. Use the rule set obtained in step S21 as the node set. Construct edge sets based on the spatial entity coordinates (longitude, latitude, elevation) and timestamps in the rule premises and conclusions. .

[0058] Spatial edge weights are calculated based on Euclidean distance: ,in For rules and The three-dimensional Euclidean distance (unit: m) involving spatial entities. The spatial attenuation coefficient (unit: m) is determined by the site inhomogeneity, typically taking a value of 0.5 times the characteristic length of the site; the time edge weight is calculated based on the time interval. ,in The time interval between rules (unit: seconds). The time decay factor (unit: s) is determined by the duration of the ground motion, and its typical value is 0.3 times the effective duration of the ground motion.

[0059] Step S22-2: Input the regular spatiotemporal graph into the spatiotemporal graph attention network for weight calculation. The network includes... The graph uses a multi-head attention mechanism in each layer. For the first layer... Layer, First Head attention, node To the neighbors The attention coefficient is calculated as follows:

[0060] .

[0061] in, For nodes In the The feature vector of the layer (dimension: ), For learnable linear transformation matrix (dimension: ), Attention parameter vector (dimension: ), This represents vector concatenation, with LeakyReLU as the activation function. Node features are updated after attention-weighted aggregation: .

[0062] Where H is the number of attention heads (typically 4 or 8). For nodes The neighborhood group, It is the activation function for the exponential linear unit.

[0063] Step S22-3: Extract output layer node features and calculate spatiotemporal correlation weights. Then, extract the output features of the last layer. Mapped to scalar weight values ​​via a fully connected layer:

[0064] The MLP contains two hidden layers, each with 64 neurons, and the activation function is a linear rectified function. That is, rules The spatiotemporal correlation weights have a value range of [0,1].

[0065] Step S22-4: Screen key disaster chains based on spatiotemporal correlation weights. Set weight thresholds. Select The rules constitute a subset of key rules; within the subset of key rules, a topological sort is performed based on chronological order and spatial adjacency to construct a directed acyclic graph, where the paths are the critical disaster chains.

[0066] For example, regular sequences The spatiotemporal correlation weights are 0.85, 0.82, and 0.79, respectively, all exceeding the threshold and satisfying the spatiotemporal adjacency condition, thus forming a critical disaster chain.

[0067] Step S23: The weighted rules are converted into a hierarchical behavioral strategy library, which is divided into three levels according to decision-making authority: high-level decision-making, mid-level allocation, and low-level control. Among them, the high-level decision-making strategy is executed by the environmental intelligent agent and is used to coordinate the earthquake propagation and disaster chain evolution; the mid-level allocation strategy is executed by the facility intelligent agent and is used to manage emergency resource scheduling and facility function switching; the low-level control strategy is executed by the structural intelligent agent and the personnel intelligent agent, respectively simulating structural damage response and personnel emergency actions. Structural damage response includes the determination of structural reinforcement areas, and personnel emergency actions include personnel evacuation route planning. The structural intelligent agent and the personnel intelligent agent execute a coordinated response according to the weighted rules.

[0068] The hierarchical behavioral strategy library refers to a set of agent behavior rules organized hierarchically according to decision-making authority. In this application, it specifically refers to a knowledge base structure that maps weighted disaster rules into collaborative strategies for four levels of agents: environment, facilities, structure, and personnel. High-level decision-making refers to the strategy level that coordinates overall disaster response objectives. In this application, it specifically refers to the seismic propagation control and disaster chain evolution guidance strategies executed by the environmental agent. Mid-level allocation refers to the strategy level that coordinates resources and functions. In this application, it specifically refers to the emergency resource scheduling and facility function switching strategies executed by the facility agent. Low-level control refers to the strategy level that executes specific physical actions. In this application, it specifically refers to the determination of structural reinforcement areas executed by the structural agent and the planning of personnel evacuation routes executed by the personnel agent.

[0069] The necessary procedures are as follows:

[0070] Step S23-1: Based on the action type and target in the rule conclusion, map the weighted rule to the corresponding strategy level. If the rule action involves global ground motion field adjustment or disaster chain phase transition (such as "entering the strong earthquake duration phase"), it is mapped to a high-level decision-making strategy; if the rule action involves resource scheduling or function switching (such as "activating the backup ventilation system"), it is mapped to a mid-level allocation strategy; if the rule action involves structural physical response or personnel movement (such as "closing a section" or "moving to the exit"), it is mapped to a low-level control strategy.

[0071] Step S23-2 involves formally encoding the high-level decision-making strategy. The high-level strategy of the environmental agent is encoded as a triple. The stages are markers of the disaster evolution phases (e.g., P1: initial earthquake, P2: strong earthquake, P3: aftershock), the objective is the global response objective function (e.g., minimizing the mean global structural damage index), and the constraints are a set of physical constraints (e.g., the time history boundary of ground acceleration). For example, the rule "If the peak ground acceleration is greater than 0.20g (approximately 1.96 m / s²), then enter the strong earthquake duration stage" is encoded as a high-rise strategy. ,in For the first Damage index of each structural unit This represents the total number of structural units. Input the time history of ground motion acceleration (unit: m / s²).

[0072] Step S23-3: Formalize the mid-level allocation strategy. The mid-level strategy of the facility agent is encoded as a quadruple. The resource is the schedulable resource type (such as ventilation volume, lighting power, drainage pump capacity), the action type is the scheduling action (such as increase, switch, shut down), the priority is the priority level (dimensionless integer, 1–10), and the condition is the trigger condition. For example, the rule "If the main ventilation system fails, then start the backup ventilation system" is encoded as a mid-level strategy (ventilation volume, increase, 8, main ventilation system status = failed).

[0073] Step S23-4: Formalize the underlying control strategy and establish linkage response relationships. The underlying strategy of the structural agent is encoded as a quintuple (position, response, intensity, duration, safety), where position is the coordinates of the center of the reinforced area (unit: m), response is the response type (e.g., shotcrete, steel support erection), intensity is the response intensity (value [0,1]), duration is the response duration (unit: s), and safety is the safety redundancy coefficient. The underlying strategy of the personnel agent is encoded as a quadruple (start point, end point, path, speed, mode), where start point is the initial position coordinates (unit: m), end point is the target safety area coordinates (unit: m), path is the evacuation path node sequence, speed is the movement speed (unit: m / s), and mode is the movement mode (0: walking, 1: running, 2: mutual assistance). The structural agent and the human agent execute a coordinated response based on weighted rules: when the structural agent determines the reinforcement area, it broadcasts the area blockade information to the human agent; after receiving the information, the human agent marks the area as an obstacle in the evacuation path planning and re-finds the route.

[0074] Step S24: Based on the hierarchical behavior policy library, parallel interaction between agents is executed through hierarchical multi-agent reinforcement learning; wherein, the high-level policy of the environment agent drives the mid-level allocation of the facility agent, the mid-level allocation of the facility agent coordinates the low-level control of the structure agent and the personnel agent, and the state of each agent is synchronized in real time through a preset communication mechanism.

[0075] Among them, hierarchical multi-agent reinforcement learning refers to a reinforcement learning framework that decomposes multi-agent decision-making problems hierarchically, with upper-level policies guiding the execution of lower-level policies. In this application, it specifically refers to an algorithm system with a three-level policy network (high-level, mid-level, and low-level) as its core, achieving collaborative decision-making through a shared state space. Parallel interaction refers to a simulation mode in which multiple agents simultaneously execute decisions and actions within the same time step. In this application, it specifically refers to a distributed computing process in which four types of agents—environment, facilities, structure, and personnel—coordinate and advance according to hierarchical permissions.

[0076] The specific process described above can be found in steps S241 to S244.

[0077] Step S25: During the parallel interaction process, a pre-set game theory negotiation mechanism is introduced to resolve the goal conflict between agents; the environmental agent acts as the coordination center, and in response to the spatial competition between personnel evacuation routes and structural reinforcement areas, an equilibrium strategy combination is obtained by solving the Nash equilibrium. The environmental agent then coordinates the conflict and optimizes the strategy combination based on the high-level decision-making strategy to generate a collaborative strategy.

[0078] Among them, the game theory negotiation mechanism refers to a decision-making mechanism based on non-cooperative game theory, which coordinates the conflicting interests of multiple agents through strategy combination and equilibrium selection. In this application, it specifically refers to a conflict resolution framework with Nash equilibrium as the solution concept and environmental agent as the coordination center. Goal conflict refers to the mutually exclusive optimization goals of different agents due to resource competition or space occupation. In this application, it specifically refers to the spatial competition between personnel evacuation routes and structural reinforcement areas. Nash equilibrium refers to the strategy combination state in non-cooperative games where each participant cannot obtain higher returns by unilaterally changing their strategy. In this application, it specifically refers to a stable configuration in terms of spatial occupation where the set of personnel evacuation routes and the set of structural reinforcement areas do not interfere with each other.

[0079] The necessary procedures are as follows:

[0080] Step S25-1: Detect target conflicts between agents. At each simulation time step, compare the set of evacuation path nodes planned by the human agents. The set of nodes in the reinforced region determined by the structural agent Computational space intersection .like If so, a target conflict is determined to exist, and the conflict intensity is quantified as follows: ,in This indicates the number of elements in the set.

[0081] Step S25-2: Construct a non-cooperative game model. Model the goal conflict as a two-player non-cooperative game. ,in For the participants to gather, For the strategy space ( For a set of variations of evacuation routes, (a set of variants for structural reinforcement regions). This is a set of profit functions. The profit function of a human agent is defined as follows: ,in Evacuation time (unit: seconds). Evacuation time is for reference (unit: s, typical value is 300s). The path hazard level is dimensionless and calculated by weighted average of the structural damage index along the path. Here, the risk weight coefficient is (dimensionless, typically 2.0); the structural agent's reward function is defined as... ,in Construction time for reinforcement (unit: seconds). For reference reinforcement time (unit: s, typical value is 600s). The intensity of ground motion during the construction period (unit: m / s²). Reference ground motion acceleration (unit: m / s², typical value is 9.8 m / s²). This is the earthquake risk weighting coefficient (typically 1.5).

[0082] Step S25-3: Solve for the Nash equilibrium policy combination. A dynamic iterative solution using the optimal response is employed: in each iteration, one policy is fixed, and the optimal response policy that maximizes the payoff of the other is found; this is repeated until both policies no longer change or the maximum number of iterations (e.g., 100) is reached. For example, the initial policy of the human agent is a path... (From area A to the exit), the initial strategy of the structural agent is to reinforce the area. (Coverage Area A); Personnel intelligent agents targeting The best response is a detour path (From Zone B to the exit), the structural intelligent agent targets The best response is to narrow the area (Only covers the edge of region A); converges to the equilibrium strategy after 5 iterations. ,in Avoid Furthermore, neither party's gains can be increased through unilateral changes.

[0083] Step S25-4 involves the environmental agent coordinating conflicts and optimizing the equilibrium strategy combination based on the high-level decision-making strategy. The environmental agent evaluates the global impact of the equilibrium strategy: calculating the global performance loss. ,in For structural damage loss, For the loss of facility function, For the risk of casualties and losses, , , These are the global weight coefficients, and + + =1. If Exceeding the preset threshold If the environmental agent initiates conflict coordination, it will forcibly adjust the reinforcement area of ​​the structural agent to the sub-area with the least impact on evacuation, or forcibly adjust the evacuation path of the personnel agent to the channel with the lowest priority of structural reinforcement, generate a coordination strategy and issue it to each agent for execution.

[0084] The execution of hierarchical multi-agent reinforcement learning requires clearly defining the decision-making authority, network architecture, communication mechanism, and parameter optimization path for each level of agent. Existing reinforcement learning methods are mostly used in single-agent or homogeneous multi-agent scenarios, making it difficult to handle the heterogeneity and hierarchy of environment, facilities, structure, and personnel in the seismic response of underground transportation engineering. This application constructs a hierarchical policy network, a lightweight message queue, and a policy gradient optimization framework to achieve gradient propagation from high-level objectives to mid-level coordination to low-level execution. The following sections disclose each sub-step.

[0085] Based on a hierarchical behavior policy library, parallel interactions between agents are executed through hierarchical multi-agent reinforcement learning, including:

[0086] In step S241, the environmental agent executes the high-level decision-making strategy and issues a mid-level allocation instruction to the facility agent. The facility agent coordinates the structural agent and the personnel agent to execute the low-level control strategy according to the instruction, and executes the structural damage response and personnel emergency actions respectively. The interaction permissions between the agents are determined according to the three levels.

[0087] Among them, the high-level decision-making strategy refers to the top-level strategy for the environmental agent to coordinate the overall earthquake response, and in this application, it specifically refers to the global target setting and constraint issuance strategy based on the judgment of the disaster chain evolution stage. The mid-level allocation command refers to the resource scheduling and function switching command generated by the facility agent after receiving the high-level strategy, and in this application, it specifically refers to the structured command containing the target facility identifier, action type, execution priority, and safety constraints. Interaction permission refers to the command transmission and response permission between agents at different levels, and in this application, it specifically refers to the unidirectional command chain of the environmental agent → facility agent → structure / personnel agent and the negotiation permission between agents at the same level.

[0088] The necessary procedures are as follows:

[0089] Step S241-1: Determine the interaction permissions between agents based on three levels, and establish a permission matrix and instruction verification mechanism. The decision-making permissions in the multi-agent system are divided into three levels: high-level, mid-level, and low-level. High-level permissions belong to the environmental agent, possessing the right to read the global state, determine the disaster stage, set global goals, and issue mid-level allocation instructions to the facility agent. Mid-level permissions belong to the facility agent, possessing the right to parse instructions, schedule resources, switch functions, and coordinate the issuance of low-level control commands to the structural and personnel agents. Low-level permissions belong to the structural and personnel agents, possessing the right to perceive local states, execute actions, and report state data to the next higher level. Agents at the same level (structural and personnel agents) have bidirectional negotiation permissions, but in cases of major conflicts such as spatial competition, arbitration by the environmental agent is required. The permission matrix is ​​stored in a two-dimensional table, with rows representing the instruction issuer and columns representing the instruction receiver. Matrix elements take values ​​{0, 1, 2}, where 0 represents no permission, 1 represents one-way instruction permission, and 2 represents bidirectional negotiation permission. For example, the configuration is as follows: Environmental agent → Facility agent = 1 (one-way command); Facility agent → Structural agent = 1 (one-way coordination); Facility agent → Personnel agent = 1 (one-way coordination); Structural agent → Personnel agent = 2 (two-way negotiation); Structural agent → Environmental agent = 0 (no bypassing of levels; must be forwarded through the facility agent or asynchronously fed back via the event bus). Before each instruction is transmitted, the sender must check the permission matrix to verify whether the communication is within the authorized scope. If the authorization is exceeded, the transmission is rejected and the anomaly is recorded.

[0090] Step S241-2: The environmental agent reads the current simulation time step under the constraints of the permission matrix. global state Including: seismic intensity index (Unit: m / s²) Set of currently active nodes in the disaster chain (Dimensionless index set), mean of global structural damage index Based on the matching rules of the current disaster phase in the high-level strategy library, the global response target is determined. For example, during the duration of a strong earthquake (Phase = P2), a global objective is set. That is, minimizing the global average damage exponent at the next time step.

[0091] Step S241-3, the environmental agent, based on the higher-level permissions, sets the global target... Decomposed into a set of facility-level sub-objectives ,in The number of facility agents. Decomposition is based on facility function weights: ,in, For the first The functional weights of each facility agent (dimensionless, satisfying...) The importance of the facility system to the overall operation is assigned (e.g., ventilation system weight 0.35, lighting system weight 0.20, communication system weight 0.25, drainage system weight 0.20). Sub-objectives are encapsulated into mid-level dispatch instructions using a pre-defined instruction encoding protocol. The instruction format is as follows: The target identifier is the target facility identifier (string), the action type is the action type enumeration value (e.g., start standby, reduce load, isolate section), the parameter list is the parameter set (e.g., {ventilation volume: 1200m³ / h, switching time: 30s}), and the timestamp is the timestamp (unit: s). After encapsulation, the environmental agent queries the permission matrix to confirm that it has one-way command permission to the target facility agent (matrix element = 1), and issues the command after the verification is successful.

[0092] In step S241-4, after receiving the mid-level dispatch instruction, the facility agent first verifies the sender's permissions (confirming it comes from the environment agent and that the matrix element = 1), and then parses the sub-targets in the instruction. The system generates coordination commands for structural and personnel agents based on the underlying control strategy library. For example, when the facility agent receives the instruction "Start the backup ventilation system, ventilation volume 1200 m³ / h", it parses it into the sub-objective "Maintain ventilation function not less than 80%"; then, based on the mid-level coordination permissions, it sends the command "Prioritize the structural stability of the ventilation room" to the structural agent and the command "Guide personnel away from the ventilation room construction area" to the personnel agent. Before sending, the permission matrix is ​​queried again to confirm that both the structural and personnel agents have one-way coordination permissions (matrix element = 1).

[0093] In steps S241-5, after receiving the coordination command, the structural agent and the personnel agent verify the sender's permissions (confirming it originates from the facility agent) and then execute the underlying control strategy according to the received coordination command. When the structural agent determines the reinforcement area, it adds "prioritizing the ventilation room" as a spatial constraint to the optimization model; when the personnel agent plans the evacuation route, it adds "away from the ventilation room construction area" as an obstacle constraint to the path planning map. If the structural agent and the personnel agent discover conflicts such as spatial competition during execution, they conduct peer-to-peer negotiation based on the two-way negotiation permissions (matrix element = 2); if the negotiation fails, the conflict is submitted to the environmental agent for arbitration according to the permission matrix.

[0094] Step S242: Based on interactive permissions, construct a hierarchical reinforcement learning architecture corresponding to the hierarchical behavior policy library. The hierarchical reinforcement learning architecture includes: a high-level policy network for executing high-level decision-making policies; a mid-level coordination network for executing mid-level allocation policies; and a low-level execution network for executing low-level control policies. The high-level policy network outputs a global disaster response objective, the mid-level coordination network breaks down the global disaster response objective into resource allocation sub-objectives, and the low-level execution network executes specific actions based on the sub-objectives, including: the structural agent calculating quantitative indicators of structural damage response, including a structural damage index; and the personnel agent executing personnel emergency actions, including planning personnel evacuation routes. Each layer of the network makes collaborative decisions through a shared state space.

[0095] In this application, the high-level policy network refers to a deep neural network used to generate global decision-making strategies for environmental agents. Specifically, it refers to a multilayer perceptron or recurrent neural network that takes the global state as input and the disaster stage and global objective as output. The mid-level coordination network refers to a deep neural network used to decompose the global objective into facility-level sub-objectives and coordinate resource allocation among facility agents. Specifically, it refers to an attention-enhanced network that takes the global objective and facility state as input and the facility action probability distribution as output. The low-level execution network refers to a deep neural network used to generate specific action instructions for structural agents and personnel agents. Specifically, it refers to a policy network that takes the local observation state and superior instructions as input and the structural response action or personnel movement action as output. The global disaster response objective refers to the disaster response optimization direction expressed mathematically. Specifically, it refers to the function that minimizes the service performance loss of underground transportation engineering in this application. The resource allocation sub-objective refers to a local optimization objective for a specific facility system obtained from the decomposition of the global objective. Specifically, it refers to the function that maximizes the facility function maintenance rate or minimizes resource consumption in this application.

[0096] The necessary procedures are as follows:

[0097] Step S242-1: Construct a high-level policy network.

[0098] The network input is a global state vector. ,in The intensity of ground motion (unit: m / s²). The average structural damage index is the global average. The average availability of facilities across the entire region. The percentage of people trapped is represented. The network structure is a two-layer Long Short-Term Memory network with 128 hidden layers. The output is the probability distribution of the disaster stage. ,in This represents the final hidden state of the Long Short-Term Memory network. To output the weight matrix, This is the bias vector. Based on... Select the current disaster phase and output the corresponding global objective. .

[0099] Step S242-2: Construct the intermediate-level coordination network. The network input is the global objective. With facility state vector ,in For the first The availability of each facility's functions For the first Each facility has available resource reserves (unit: based on resource type, such as kW, m³ / h, etc.). The network employs an attention mechanism.

[0100] First, calculate the inter-facility attention weights. ,in , , These are learnable parameters; the Xavier initialization method is used to sample from a zero-mean Gaussian distribution with a standard deviation of . ,in, , These are the input and output neuron dimensions, determined by the number of facilities and the state vector dimension.

[0101] Then, the context vector is obtained by weighted aggregation of the states. Finally, the probability distribution of each facility's actions is output through two fully connected layers (64 hidden dimensions, linear rectified activation function). For example, for ventilation and lighting systems, the probabilities of the output actions are respectively... (Corresponding to {maintain, reduce load, switch to standby}) and .

[0102] Step S242-3: Construct the underlying execution network. The network is divided into a structural execution subnetwork and a personnel execution subnetwork.

[0103] The input to the structure execution subnetwork is the local structure state. ,in This is the local damage index. This represents the local maximum stress (unit: MPa). This represents the local maximum displacement (unit: mm). Encodes coordination commands (dimensionless integers) from facility agents. The network consists of 3 fully connected layers (hidden dimensions 128-64-32, activation function with leaky linear rectified), and the output is the structure response action. This includes the reinforcement location coordinates (unit: m), reinforcement strength (value [0,1]), and reinforcement duration (unit: s). The personnel execution subnetwork input is the local personnel state. ,in The coordinates of the personnel's location (unit: m). The speed is represented by m / s, and panic index is represented by a value of [0,1]. The network consists of three fully connected layers (64-32-16 hidden dimensions, hyperbolic tangent activation function), and the output is the movement motion of the person. This includes the target direction angle (unit: rad), movement speed (unit: m / s), and movement mode (dimensionless integer, 0: walking, 1: running, 2: mutual assistance).

[0104] Step S242-4: Establish a shared state space collaborative decision-making mechanism. Key indicators in the global state are shared across network layers: the disaster phase output by the high-level strategy network serves as the conditional input for the mid-level coordination network; the facility actions output by the mid-level coordination network serve as the instruction input for the bottom-level execution network; and the structural damage index output by the bottom-level execution network is used to inversely update the global state based on personnel location coordinates. and The shared state space vector has a dimension of 256 and is uniformly mapped through a pre-defined state encoder. The encoder consists of two fully connected layers (128 hidden dimensions and ReLU activation function).

[0105] Step S243: Transmit the agent status data generated by collaborative decision-making in real time through a preset lightweight message queue, and store the status data and agent actions in the interaction log. The status data includes the structural damage index calculated by the structural agent, the facility resource reserve monitored by the facility agent, and the personnel location coordinates obtained by the personnel agent.

[0106] Lightweight message queues refer to inter-process communication middleware characterized by low latency and high throughput. In this application, they specifically refer to publish-subscribe message buses used for broadcasting state data among multiple agents, such as Zero Message Queue (ZMQ) or the topic mechanism of Robot Operating System (ROS). Interaction logs refer to time-series databases that record state data and action commands throughout the multi-agent simulation process. In this application, they specifically refer to structured log files indexed by time steps that support fast retrieval and replay.

[0107] The necessary procedures are as follows:

[0108] Step S243-1: Deploy a lightweight message queue communication layer. A publish-subscribe pattern is used to construct the communication topology: the environment agent acts as the publisher, publishing disaster stage and global target messages to the topic "Global Stage"; the facility agent subscribes to "Global Stage" and acts as the publisher, publishing facility coordination commands to the topic "Facility Commands"; the structure agent and personnel agent subscribe to "Facility Commands" and act as publishers, publishing local states to the topics "Structure Status" and "Personnel Status" respectively. Message serialization uses Protocol Buffers (Protobuf) format, and the message body includes a timestamp (unit: s), agent identifier (string), state vector (floating-point array), and action vector (floating-point array).

[0109] Step S243-2: Define the status data format and transmission period. Status data includes: structural damage index. (Calculated by the structural intelligent agent) Facility resource surplus (Units: based on resource type, such as kWh, m³, etc., obtained from facility intelligence monitoring), personnel location coordinates (Unit: m, obtained from human agent positioning). Transmission period With simulation time step Consistent, typically taking a value of 0.1s, to ensure real-time synchronization of status data.

[0110] Step S243-3: Store the state data and agent actions in the interaction log by time step. The interaction log adopts a hierarchical storage structure: the first layer is the time step index; the second layer is the agent type identifier (enumerated values: environment / facilities / structures / personnel); the third layer is the data record, in the format of (timestamp, agent identifier, state vector, action vector, reward signal).

[0111] For example, at time step (Corresponding to simulation time 10.0s), the interaction log of the structural agent STR-03 is recorded as follows: timestamp 10.0s, agent identifier STR-03, state vector [0.35, 45.2, 12.8, 0], action vector {reinforcement region center coordinates (120.5, 30.2, −15.0) m, reinforcement strength 0.6, duration 300.0s}, reward signal -2.5. The components of the state vector are: local damage index 0.35, maximum stress 45.2MPa, maximum displacement 12.8mm, and coordination command code 0 (indicating no coordination command). The action vector represents implementing structural reinforcement with a strength of 0.6 for 300.0s at coordinates (120.5, 30.2, −15.0) m. The reward signal -2.5 represents the performance penalty caused by this action.

[0112] Step S244: With the optimization objective of minimizing the service performance loss of underground transportation engineering, including structural damage and functional failure, the network parameters of each layer are dynamically adjusted based on the agent's actions and state data recorded in the interaction log through a preset policy gradient algorithm.

[0113] In this context, the policy gradient algorithm refers to a reinforcement learning optimization method that directly calculates the gradient of the policy network parameters with the goal of maximizing the expected cumulative reward. In this application, it specifically refers to the Proximal Policy Optimization (PPO) algorithm or its hierarchical variants, used to update the weights of high-level, mid-level, and low-level networks. Network parameters refer to the set of learnable weight matrices and bias vectors in a neural network; in this application, they specifically refer to the tensors updated via backpropagation in each layer of the policy network.

[0114] The necessary procedures are as follows:

[0115] Step S244-1, define the service performance loss function. (Comprehensive loss function) It is composed of a weighted average of structural damage loss, facility function loss, and loss of personnel injury risk:

[0116] .

[0117] in, , , The loss weight coefficients (satisfying) Typical values ); , is the average damage index of all structural units. This represents the total number of structural units. This represents the average percentage of all facility malfunctions. For the first The availability of each facility's functions Total number of facilities; , representing the average loss of safety status for all personnel. For the first Safety status index of individuals (1 indicates safety, 0 indicates being trapped / injured). This refers to the total number of personnel.

[0118] Step S244-2: Extract training samples from the interaction log. Extract state-action-reward triples by time step. Among them, the reward .in, The total loss at the current time step, This is the penalty coefficient for the action (typically 0.5). It is the L2 norm of the agent's action vector (normalized by the maximum action amplitude). The survival reward constant (typically 2.0). Analysis process: The reward function consists of three terms: (1) The greater the loss, the more negative the reward, driving the strategy to reduce damage; (2) Action penalty: Suppresses meaningless zero actions—the agent must take effective actions to reduce [the penalty]. However, excessive actions will be punished, prompting the strategy to choose concise and efficient emergency actions; (3) : Baseline offset, ensuring performance under extreme and harsh conditions ( And if the action is significant, the reward will not be too negative, maintaining numerical stability. If the agent takes no action at all, ,but It may be due to the natural growth of disaster evolution, leading to Decrease; conversely, taking effective actions can reduce. Despite the penalty of actions, the net reward may still increase. This prevents the policy from degenerating into zero actions. Experience replay buffers are constructed for the high-level, mid-level, and low-level networks, with a buffer capacity of [missing information]. Each record is subject to a first-in, first-out (FIFO) elimination mechanism.

[0119] Step S244-3: Perform hierarchical policy gradient update. For the high-level policy network, the PPO-Clip objective function is used:

[0120] .

[0121] in, For strategy ratio, This is the estimate of the generalized advantage. This is the cropping factor (typically 0.2). The pruning function is used. A PPO-Clip structure is adopted for the mid-level coordination network and the low-level execution network, but the advantage functions are calculated based on facility-level rewards and local rewards, respectively. The optimizer uses Adam, and the learning rate is... The batch size is 64, and 2048 time steps of data are collected in each iteration. A total of 10 rounds of mini-batch gradient descent are performed.

[0122] Step S244-4: Perform network parameter synchronization and convergence determination. After every 5 rounds of policy updates, synchronize the current network parameters to the policy network replicas of each agent; calculate the average reward change rate over the last 1000 steps. ,like If the decision is made three times consecutively, the decision strategy converges, and training stops; otherwise, iteration continues. For example, after 500 rounds of updates, the average reward increases from the initial value of -45.2 to -12.8, and the rate of change decreases to 8 × 10⁻⁶. -5 Once convergence is determined, the network parameters are fixed for subsequent simulations.

[0123] The bidirectional coupling mechanism for establishing knowledge graphs and multi-agent simulation models includes:

[0124] Step S31: Extract multi-dimensional features of semantic rules from the knowledge graph using a preset graph neural network algorithm, and convert the multi-dimensional features into initialization parameters and behavioral constraints of the multi-agent simulation model; wherein, the graph neural network algorithm integrates an attention mechanism to prioritize the processing of disaster chains reflected by semantic rules.

[0125] Graph neural network algorithms refer to neural network algorithms that directly perform message passing and feature aggregation on graph-structured data. In this application, they specifically refer to graph representation learning algorithms that take knowledge graphs as input and node / edge embedding vectors as output. Multidimensional features refer to vector representations that contain rich semantic information after being encoded by graph neural networks. In this application, they specifically refer to low-dimensional dense vectors that integrate node attributes, neighborhood structure, and relationship types, with typical dimensional values ​​of 128 or 256.

[0126] The necessary procedures are as follows:

[0127] Step S31-1: Construct the graph neural network input representation of the knowledge graph. Represent the knowledge graph as a heterogeneous graph. ,in It is a set of nodes (including four types of nodes: disaster-bearing bodies, disaster-causing factors, disaster-prone environments, and personnel). Let be the set of edges. For node type mapping functions, This is a set of relation types. For each node type... Initialize learnable type embeddings ,in For type embedding dimension (typically 32); for each edge relationship Initialize learnable relation embeddings ,in This represents the relation embedding dimension (typically 32). Initial node features. The following are obtained from instance attribute encoding: numerical attributes (such as lining thickness 0.35m, peak acceleration 0.20g) are normalized and then spliced; categorical attributes (such as site category = hard soil) are uniquely thermally encoded and then spliced.

[0128] Step S31-2: Perform graph neural network message passing and feature aggregation. A Relational Graph Convolutional Network (R-GCN) is used for this process. Layer message passing. In the first layer... Layer, for nodes Type-specific aggregations are calculated as follows:

[0129] .

[0130] in, For nodes In the The feature vector of the layer (dimension: ), For nodes In relationship The following is a collection of neighbors. The normalization constant (take) ), For relationship In the Learnable weight matrix of layer (dimension: ), Self-connection weight matrix (dimension: ), This is the ReLU activation function. After layer propagation, the final node embedding is obtained. Dimension .

[0131] Step S31-3: Embed the node mapping to initialization parameters. For disaster-bearing nodes (structural classes), embed them and map them through a fully connected layer to structural mechanical parameters: elastic modulus. (Unit: GPa, where) The baseline elastic modulus is determined by the material specifications (e.g., 30 GPa for concrete); density. (Unit: kg / m³). For facility nodes, this is mapped to facility status parameters: initial functional availability. (Values ​​[0,1]); Initial resource reserves (Unit: based on resource type). For personnel nodes, the mapping is to personnel distribution parameters: initial location coordinates. (Unit: m); Initial Panic Index (Values ​​[0,1]).

[0132] Step S31-4: Map edge embeddings and path features to behavioral constraints. For semantic relation edges... Extract its relation embedding Edge features are obtained by concatenating the embeddings of the two endpoints. The constraint decoder outputs behavioral constraints. These constraints include: physical constraints (such as the maximum allowable stress of the structure). (Unit: MPa) and resource constraints (such as minimum resource margin of facilities) (Unit: based on resource type) and safety constraints (such as maximum permissible exposure time for personnel) (Unit: s). For example, decoding the "acting on" relationship edge between "historical earthquake event M" and "left tunnel of section XX" yields the physical constraints. MPa (corresponding to the concrete compressive strength grade) represents the maximum stress that the structure can withstand under seismic loading.

[0133] Step S32: Based on the event-driven architecture, collect state change data generated by each agent during the operation of the multi-agent simulation model in real time; the state change data includes structural damage index, facility resource reserves and personnel location coordinates; process the state change data through an incremental learning algorithm controlled by a time window, dynamically update the attribute values ​​of instances corresponding to the ontology in the knowledge graph, and adjust the association strength of semantic relationships between instances; the attribute values ​​and association strength are based on the ontology definition.

[0134] Here, a time window refers to the time interval for discretizing a continuous time series. In this application, it specifically refers to a fixed duration or fixed step size interval used for batch processing of state change data, typically 5 seconds or 10 simulation steps. An incremental learning algorithm refers to a learning algorithm that continuously updates model parameters based on newly arriving data without forgetting existing knowledge. In this application, it specifically refers to an online learning mechanism used to dynamically update the attribute values ​​and association strengths of a knowledge graph, such as online gradient descent or Bayesian update.

[0135] The necessary procedures are as follows:

[0136] Step S32-1: Define the event-triggered acquisition mechanism for state change data. Set the triggering conditions for three types of core events: structural damage event—when the structural unit damage index... Exceeding the previous time step and increment Time-triggered; Facility degradation event—when facility functionality availability is reduced. Decline and the magnitude of the decline Time-triggered; Personnel movement event—when the rate of change of personnel position coordinates m / s and the direction deviates from the reference path angle Events are triggered on demand. Once an event is triggered, the corresponding agent immediately publishes a state snapshot to the event bus.

[0137] Step S32-2: Aggregate state change data through the event bus and divide it into blocks according to time windows. The event bus uses a sliding time window mechanism: window length... s, sliding step size Within each window, aggregate state snapshots triggered by all events during that time period to form a data block. Each snapshot contains a timestamp, agent identifier, state type, state value, and change amount.

[0138] Step S32-3: Perform incremental feature extraction on the data block. For structural damage data, extract the feature: damage index increment. Damage propagation rate (in (Window duration); for facility data, extract features: rate of change in functional availability. (in (For the previous window's functionality availability); for personnel data, extract the feature: path deviation. (in For actual location, For the planned location, (This refers to the planned path length).

[0139] Step S32-4: Update the attribute values ​​of knowledge graph instances based on the incremental learning algorithm. Online gradient descent is used to update numerical attributes: for the structural damage index attribute, the update formula is as follows... ,in This is the learning rate (typically 0.1). For the damage propagation rate, The window duration (in seconds) ensures that the damage exponent monotonically increases over time with an upper limit of 1.0. Boundary conditions: If Then let ;like (Caused by numerical noise), then let Because of the updated formula , Theoretically, it is monotonically non-decreasing and will not naturally fall below 0. However, to mitigate floating-point rounding errors, a lower limit truncation is retained. An upper limit truncation ensures the damage index does not exceed 1.0 (complete damage). The update formula for the facility availability attribute is as follows: ,in The vulnerability factor is determined by the type of facility, such as 0.3 for ventilation systems and 0.2 for lighting systems.

[0140] Step S32-5: Update the semantic relationship strength based on the incremental learning algorithm. For the semantic relationship edges involved in the triggering event, use attention-enhanced Bayesian updating to obtain the new relationship strength. ,in, To observe the correlation strength (calculated from the event co-occurrence frequency): , For nodes and Number of times co-occurring within the window (Total number of events within the window) This is the update rate coefficient (typically 0.3). The graph attention coefficients (obtained from the graph neural network attention weight transfer in step S31). For example, in the window... Within the structure node "Left Line of Tunnel in Section XX" and the facility node "Main Ventilation System" appear 3 times, for a total of 10 events. If the old correlation strength , , Then the new correlation strength This indicates that the relationship between the two within the window is weakened due to insufficient co-occurrence of events.

[0141] Step S33: Complete the bidirectional coupling between the knowledge graph and the multi-agent simulation model.

[0142] Among them, bidirectional coupling refers to the technical state in which a data closed loop is formed between the knowledge graph and the multi-agent simulation model. In this application, it specifically refers to the state in which the complete path of semantic rules driving simulation in the forward direction and the simulation state updating the knowledge graph in the reverse direction has been established and verified.

[0143] The necessary procedures are as follows:

[0144] Step S33-1, Verify the positive coupling path: The semantic rules extracted from the knowledge graph are successfully converted into the initialization parameters and behavioral constraints of the multi-agent simulation model after being encoded by the graph neural network, and the simulation model is in the initial time step The system can start normally and load the above parameters. Verification method: Check whether the initial state of each agent in the simulation log is consistent with the corresponding instance attributes in the knowledge graph, such as whether the initial elastic modulus of the structural agent is equal to the elastic modulus attribute value of the "lining concrete" instance in the knowledge graph.

[0145] Step S33-2, Verify the reverse coupling path: The state change data generated during the operation of the multi-agent simulation model, after being collected and incrementally learned by the event-driven architecture, is successfully written into the attribute value field of the corresponding instance in the knowledge graph, and the association strength update value is within a reasonable range [0,1]. Verification method: Select 10 random time steps and compare the consistency between the instance attribute values ​​in the knowledge graph and the output state values ​​of the simulation model, requiring an error of less than 1%.

[0146] Step S33-3: If both the forward and reverse coupling paths pass verification, the bidirectional coupling mechanism is determined to be established successfully, and a coupling status flag is output. The coupling establishment timestamp is recorded. If any path verification fails, the process reverts to step S31 to re-execute feature extraction and parameter mapping until verification passes. For the specific execution process of real-time acquisition and incremental learning updates in the event-driven architecture, please refer to steps S3A to S3F; for the specific execution process of update validity verification, please refer to steps S3E1 to S3E4. Details are omitted here.

[0147] The system collects state change data in real time during the simulation process through a pre-defined event-driven architecture, feeds it back to the knowledge graph, and updates the association strength of instance attributes and semantic relationships, including:

[0148] Step S3A: Define preset event triggering conditions. Event triggering conditions include structural damage index exceeding preset damage threshold, personnel location coordinates deviating from preset baseline path by more than preset distance, and facility resource reserves being lower than preset resource threshold.

[0149] The necessary procedures are as follows:

[0150] Step S3A-1: Define the structural damage index triggering conditions. Set the damage threshold. When the damage index of any structural unit When this threshold is reached, a structural damage event is triggered. This threshold is defined based on the lower limit of the "severe damage" level in the seismic design code for underground transportation engineering structures, representing a critical state where the structural bearing capacity has significantly degraded and immediate reinforcement measures are required.

[0151] Step S3A-2: Define the triggering conditions for personnel position coordinate deviation. Set the reference path set. ,in For the path node coordinate sequence (unit: m); set the deviation distance threshold. m. When the actual coordinates of the personnel closest distance to the baseline path When this occurs, a personnel path deviation event is triggered. This threshold is based on the width of the personnel evacuation safety corridor and indicates that personnel have left the safe evacuation route and may have entered a danger zone.

[0152] Step S3A-3: Define the facility resource surplus trigger conditions. Set the resource threshold set. ,in For the first Minimum safety margin for facility resources (unit: based on resource type, e.g., 200 kWh for ventilation systems, 50 m³ / h for drainage systems). When facility resource margin... When this occurs, a facility resource shortage event is triggered. This threshold is determined based on the facility system design redundancy and minimum emergency operation requirements, indicating that the facility can no longer meet the basic functional maintenance needs.

[0153] Step S3A-4: Encode the above triggering conditions into the decision function of the event listener and deploy it on the monitoring node of the event-driven architecture. The listener periodically... s polls the state registers of each agent and performs Boolean decisions: .like Then, a trigger signal is sent to the event bus.

[0154] Step S3B: Collect agent state data in real time according to event triggering conditions through a preset communication mechanism, and filter it according to preset rules.

[0155] The communication mechanism refers to the set of protocols and channels for data exchange among multiple agents, and in this application, it specifically refers to a publish-subscribe communication protocol based on a lightweight message queue. The filtering rule refers to the criteria for removing redundant or invalid state data, and in this application, it specifically refers to the filtering logic based on the magnitude of change and data integrity.

[0156] The necessary procedures are as follows:

[0157] Step S3B-1: Establish a real-time data acquisition channel. The structural agent, facility agent, and personnel agent each publish status data to the designated message topic: the structural agent publishes the damage index to the topic "structural damage". and corresponding structural unit identifiers; the facility agent publishes resource availability to the theme "facility resources". (Unit: based on resource type) and facility identification; the personnel agent publishes location coordinates to the topic "Personnel Location". (Unit: m) and personnel identification. The message publishing frequency is synchronized with the simulation step size, typically 10Hz (i.e., once every 0.1s).

[0158] Step S3B-2, perform data filtering. Set the filtering rules: (1) Change amplitude filtering - if the absolute value of the difference between the current state value and the previous sample value is less than the minimum change threshold. If the data is not found, then discard it. (Regarding the damage index) (Regarding resource reserves) m (for location coordinates); (2) Integrity filtering—if the message body is missing key fields (such as timestamp, agent identifier, state value), then discard the data; (3) Timeliness filtering—if the difference between the message timestamp and the current simulation time exceeds the lag threshold. If the time stamp value is less than 0.01, the data is discarded. For example, if the damage index of a structural agent changes from 0.352 to 0.355 within a 0.1s step, and the difference is 0.003 << 0.01, the data is filtered out; if a message is missing its timestamp field due to network jitter, the message is filtered out.

[0159] Step S3C: Divide the collected agent state data into data blocks according to the preset time window for processing and extract key features; key features include deviation features of the preset baseline path.

[0160] Among them, key features refer to derived variables extracted from the original state data that can characterize the trend of disaster evolution or abnormal patterns. In this application, they specifically refer to high-order statistics or physical indicators used to drive knowledge graph updates.

[0161] The necessary procedures are as follows:

[0162] Step S3C-1: Perform time window division. A sliding window mechanism is used, with a window length of... s, sliding step size s. Perform aggregation and feature extraction on the state data set within each window.

[0163] Step S3C-2: Extract key structural damage features. This involves analyzing the structural damage index sequence within the window. Extract: (1) Mean damage (2) Damage standard deviation (3) Damage propagation acceleration (Unit: s) -2 ),in The difference in damage rates between adjacent windows. The window sliding step size is 2.5s.

[0164] Step S3C-3: Extract key features of personnel path deviation. This involves analyzing the sequence of personnel position coordinates within the window. Extract: (1) Length of deviation from trajectory (Unit: m); (2) Deviation duration (Unit: s), where The simulation step size is 0.1s; (3) Deviation from the danger level ,in For personnel location The damage index of the corresponding structural unit represents the degree of structural danger in the area where personnel are located.

[0165] Step S3C-4: Extract key features of facility resource depletion. This involves analyzing the facility resource surplus sequence within the window. Extract: (1) Resource decay rate (Unit: s⁻¹); (2) Resource depletion early warning time (Unit: s), where The rate of change of resource surplus (unit: based on resource type / s) is obtained from the data within the linear fitting window.

[0166] In step S3D, a preset incremental learning algorithm is used to update the attribute values ​​of instances corresponding to the ontology in the knowledge graph and the association strength of semantic relationships between instances based on the processed data blocks; the attribute values ​​and association strength are based on the ontology definition.

[0167] In this context, the attribute value of an instance refers to the feature value of a specific entity instance in the knowledge graph data layer. In this application, it specifically refers to dynamic attributes such as structural damage index, facility availability, and personnel location coordinates after being updated by simulation state data. Association strength refers to the weight value of semantic relationships. In this application, it specifically refers to a scalar value that quantifies the closeness of association between instances after being dynamically adjusted through incremental learning.

[0168] The necessary procedures are as follows:

[0169] Step S3D-1: Establish the mapping relationship between data blocks and knowledge graph instances. Based on the correspondence table between agent identifiers (Agent_ID) in the data block and globally unique identifiers (Instance_URI) of knowledge graph instances, map structural damage features to the attribute field "structural damage index" of the disaster-bearing body instance, map personnel deviation features to the attribute field "path deviation status" of the personnel instance, and map facility resource features to the attribute field "resource margin" of the facility instance.

[0170] Step S3D-2: Update instance attribute values. For numeric attributes, use exponential smoothing for updates: ,in These are the observed feature values ​​extracted from the data block. These are old attribute values ​​in the knowledge graph. This is a smoothing coefficient (typically 0.7, indicating that the new data has higher reliability than the old data). For example, the old damage index of the disaster-bearing entity instance "left tunnel of section XX". Observed mean of data blocks After the update .

[0171] Step S3D-3: Update the semantic relationship strength.

[0172] For semantic relationships involving instance pairs in the data block, a collaborative filtering approach is used for updates: ,in For example and Feature similarity in data blocks (such as cosine similarity or Pearson correlation coefficient). To update the confidence score (calculated as the ratio of the number of times two instances co-occur in the data block to the total number of events in the window), This represents the learning rate (typically 0.2). For example, the instances "Left Line of Tunnel in Section XX" and "Main Ventilation System" appear 5 times in the window, for a total of 20 events. Feature similarity old correlation strength ,but The correlation strength is slightly enhanced.

[0173] Step S3E involves synchronizing the updated knowledge graph content to the multi-agent simulation model and verifying the effectiveness of the update.

[0174] Synchronization refers to the process of pushing updated knowledge graph data to the multi-agent simulation model and replacing the corresponding state variables. In this application, it specifically refers to the bidirectional data replication mechanism of attribute values ​​and association strength. Update effectiveness refers to the degree to which the prediction results generated by the updated knowledge graph content in the simulation model are consistent with physical consistency. In this application, it specifically refers to the quality index determined by comparing the prediction deviations of adjacent time steps.

[0175] The necessary procedures are as follows:

[0176] Step S3E-1: Synchronize the updated knowledge graph instance attribute values ​​to the corresponding agent state register. Through a pre-defined data mapping interface, write the structural damage index of the disaster-bearing entity instance into the structural agent state variable. Write the remaining resources of the facility instance into the facility agent's state variables. Write the location coordinates of the personnel instance into the personnel agent's state variable. The synchronization period is consistent with the simulation step size to ensure that the knowledge graph data used in each simulation step is the latest version.

[0177] Step S3E-2: Synchronize the updated semantic relationship strength to the agent interaction weight matrix. Relationship Strength Attention weights for inter-agent message passing: A pre-defined mapping table maps the knowledge graph relation type "structure-facility" to agent communication weights. The mapping table consists of an ontology relation set. The intersection of the communication topology with the intelligent agent is determined to ensure consistency of dimensions across modules; when facility intelligent agents communicate with human intelligent agents, the importance weight of messages is determined. Weights are normalized to the [0,1] range to ensure that message queue priority scheduling is consistent with association strength.

[0178] Step S3E-3: Perform a single-step simulation and verify the effectiveness of the update. For details on verifying the effectiveness of the update, please refer to steps S3E1 to S3E4; they will not be elaborated here.

[0179] Step S3F involves associating the verified update results with the state changes of the multi-agent simulation model and storing them to form a closed-loop record.

[0180] Closed-loop records refer to time-series logs that record the correspondence between knowledge graph updates and simulation state changes. In this application, they specifically refer to associated data tables that support post-event traceability and model auditing.

[0181] The necessary procedures are as follows:

[0182] Step S3F-1: Construct the closed-loop record data structure. The record entry format is: (Update Identifier, Timestamp, Update Instance, Update Attribute, Old Value, New Value, Simulation State Change, Verification Result). Here, the Update Identifier is the update identifier (string), the Timestamp is the update time (unit: seconds), the Update Instance is the update instance identifier, the Update Attribute is the update attribute name, the Old Value and New Value are the values ​​before and after the update, the Simulation State Change is the change in the state of the simulation model after synchronization, and the Verification Result is the verification result flag (TRUE indicates valid, FALSE indicates invalid).

[0183] Step S3F-2: Write the closed-loop record according to the time step. For example, in At time s, the "Structural Damage Index" attribute of the disaster-bearing entity instance "Left Line of Tunnel in Section XX" is updated from 0.399 to 0.452. After being synchronized to the structural agent, the simulation model predicts the damage index of this structural unit to be 0.461 in the next time step, and the state change is... If the verification is valid (verification result = TRUE), then the record entry is (UPD-20250519-001, 25.0, XX section tunnel left line, structural damage index, 0.399, 0.452, +0.009, TRUE).

[0184] Step S3F-3: Persistently store the closed-loop records in a time-series database (such as InfluxDB or TimescaleDB), creating an index with the timestamp as the primary key and the instance identifier as the secondary key, supporting fast queries by time range or instance identifier. Set the database retention policy to automatic archiving: data from the most recent 7 days is stored in a hot partition (SSD), and historical data is migrated to a cold partition (HDD), with a storage period of no less than 5 years.

[0185] Verifying the validity of the update includes:

[0186] In step S3E1, after synchronizing the updated knowledge graph content to the multi-agent simulation model, a preset single-step simulation is performed to obtain the prediction result for the current time step.

[0187] In this context, single-step simulation refers to the numerical calculation process that uses the current time step state as the initial condition and advances by one simulation step to obtain the state of the next time step. In this application, it specifically refers to the incremental inference operation of a multi-agent simulation model after receiving an update from the knowledge graph. The prediction result refers to the estimated state value of the next frame output by the single-step simulation. In this application, it specifically refers to the set of predicted values ​​for the structural damage index, facility functional availability, and personnel location coordinates.

[0188] The necessary procedures are as follows:

[0189] Step S3E1-1: Load the synchronized knowledge graph state as the initial condition for single-step simulation. Read the current value from the state register of each agent: structural damage index. Facility availability Personnel location coordinates (Unit: m) Environmental ground motion acceleration (Unit: m / s²).

[0190] Step S3E1-2: Perform a single-step dynamics simulation. The structural agent uses the current damage index... As initial values, based on the damage evolution equation Calculate the damage index at the next time step, where The first-order rate of change of the damage index (unit: s) -1 ), The second-order rate of change of the damage index (unit: s) -2 ), The simulation step size (unit: s). (Unit: s, typical value is 0.1s). The above formula is the damage index. The numerical extrapolation of the second-order Taylor expansion at the discrete time step is not a physical constitutive equation. The stress, strain, and damage state of the structure at the current time step are determined according to the elastoplastic damage mechanics evolution law (e.g., ,in (equivalent plastic strain rate). From two adjacent time steps Difference yields ( This extrapolation formula is applicable to... For sufficiently small time intervals (e.g., 0.1s) and smooth damage evolution, when damage changes abruptly (e.g., concrete crushing), switch to single-step direct integration.

[0191] The facility agent calculates based on the state transition matrix. The human agent calculates integrals based on the social force model. .

[0192] Step S3E1-3: Summarize the single-step simulation prediction results. Output the prediction result vector. And record the simulation time. (Unit: ms) Used for performance monitoring.

[0193] Step S3E2: Compare the prediction result of the current time step with the prediction result of the previous time step, and calculate the prediction deviation.

[0194] The prediction bias refers to the difference between the prediction results of two adjacent time steps, and in this application, it specifically refers to the quantitative indicator used to determine whether the knowledge graph update introduces anomalies.

[0195] The necessary procedures are as follows:

[0196] Step S3E2-1: Extract the prediction results from the previous time step. Prediction results at the current time step The prediction result of the previous time step is read from the prediction buffer of the simulation model, and the prediction result of the current time step is read from the output of step S3E1.

[0197] Step S3E2-2: Calculate the deviation of each component. Structural damage index deviation. ,in For the previous time step Predicted values ​​at any given time; deviation in facility availability Personnel position coordinate deviation (Unit: m).

[0198] Step S3E2-3: Calculate the overall forecast deviation. Use the weighted average overall deviation: ,in , , Weighting coefficients (typical values) ), m is the deviation distance threshold defined in step S3A-2, used to measure the position deviation. Normalize to a dimensionless quantity to ensure consistency of the three dimensions.

[0199] In step S3E3, if the prediction deviation is less than or equal to the preset error threshold, the update is confirmed to be valid, and the updated knowledge graph content is used as the basis for the simulation in the next time step.

[0200] Among them, the error threshold refers to the upper limit of the acceptable range of prediction deviation. In this application, it is a special indicator quantity, which is comprehensively calibrated by the system accuracy requirements and computing resource constraints.

[0201] The necessary procedures are as follows:

[0202] Step S3E3-1: Set the error threshold. Structural damage index error threshold. Facility availability error threshold Personnel position coordinate error threshold m. Overall deviation threshold Determined by the weighted formula:

[0203] .

[0204] Step S3E3-2, perform the judgment. If Then output the verification flag. It also sends a command to the simulation control module to "continue simulation using the updated state"; at the same time, it writes the updated knowledge graph content into the basic state library of the simulation model as the initial condition for the next time step simulation.

[0205] Step S3E3-3: If the verification passes, record the verification log. The log content includes: timestamp, updated batch number, overall deviation value, deviation values ​​of each component, threshold comparison results, and verification conclusion.

[0206] In step S3E4, if the prediction deviation is greater than the preset error threshold, the parameter adjustment of the incremental learning algorithm is triggered, and the dynamic update is re-executed until the error requirement is met or the preset number of iterations is reached.

[0207] Parameter tuning refers to the operation of modifying the hyperparameters within the incremental learning algorithm to improve the update quality; in this application, it specifically refers to the re-optimization of the learning rate, smoothing coefficient, or attention weights. Iteration count refers to the maximum allowed number of times the update-verification loop is re-executed; in this application, it specifically refers to the termination condition to prevent infinite loops.

[0208] The necessary procedures are as follows:

[0209] Step S3E4-1: Determine if the update has failed and initiate parameter adjustment. If... Then output the verification flag. This triggers an incremental learning algorithm parameter adjustment strategy: adjusting the learning rate. Reduced to (as in the original) Adjust to 0.1), smoothing coefficient Reduced to (as in the original) The weight of historical data was adjusted to 0.56 to enhance the weight of new data and suppress noise in new data.

[0210] Step S3E4-2: Re-execute the dynamic update. Use the adjusted parameters. , Re-execute the incremental learning update in step S3D to obtain new attribute values ​​and association strengths; then, repeat the verification process from steps S3E1 to S3E3 to calculate the new prediction bias. .

[0211] Step S3E4-3, Iteration Termination Determination. Set the maximum number of iterations. If in the first The next iteration ( )hour If the update is successful, output the result. ;like If the threshold is still not met, the update is considered to have failed, and the following output is given: And trigger exception handling: roll back to the previous valid knowledge graph version, mark the current data block as an abnormal sample and do not update it for the time being, and send an alarm message to the system administrator.

[0212] Parallel simulation of semantic reasoning and multi-agent simulation models is performed using knowledge graphs to dynamically predict the service performance loss state of underground transportation engineering projects. The output quantitative assessment results of structural damage and functional failure include:

[0213] Step S41: Input the ground motion parameters. Based on the ontology in the knowledge graph, analyze the semantic relationships between the ground motion parameters and disaster-causing factors, disaster-prone environment, disaster-bearing bodies and people, and activate the instances corresponding to the semantic relationships.

[0214] Among them, seismic motion parameters refer to the set of physical quantities describing the characteristics of seismic motion, specifically peak ground acceleration, significant duration, site characteristic period, and input time history in this application. Activation refers to the operation of retrieving and marking relevant instances as valid instances of the current simulation from the knowledge graph based on matching conditions, specifically referring to the instance filtering process executed by a semantic reasoning engine using seismic motion parameters as query conditions in this application.

[0215] The necessary procedures are as follows:

[0216] Step S41-1: Analyze the input ground motion parameters. Input parameters include: peak ground acceleration (unit: m / s²), significant duration. (Unit: s, defined as the time interval during which Arias intensity accumulates from 5% to 95%), site characteristic period (Unit: s) Seismic ground acceleration time history (Unit: m / s²). For example, input the parameters for a design earthquake: peak ground acceleration. m / s², s, s, time-course sampling frequency 50Hz.

[0217] Step S41-2: Construct a query graph based on ontology semantic relationships. Using the concept of "disaster-causing factor" as the root node, construct the query condition: the peak acceleration of the disaster-causing factor. Input peak acceleration and duration of catastrophic factors Furthermore, the characteristic cycle of disaster-prone environments Subgraph matching is performed in the knowledge graph using a semantic reasoning engine, such as the Pellet inference engine based on description logic or a custom graph traversal algorithm.

[0218] Step S41-3: Activate the instance. For successfully matched instances, set the activation flag. The instance attributes are extracted as initial conditions for simulation. For example, if the query matches the disaster-causing factor instance "design earthquake E1", the disaster-prone environment instance "soft soil site S1", the disaster-bearing body instance "left line of tunnel in section XX" and the personnel instance "peak period personnel distribution H1", the above instances are activated and participate in subsequent simulations.

[0219] Step S42: Based on the activated instances, extract key semantic rules using a preset graph neural network algorithm to generate a dynamic prediction rule set.

[0220] Among them, the dynamic prediction rule set refers to the disaster rule subset that is updated as the simulation progresses and is used to drive the interaction of multiple agents. In this application, it specifically refers to the rule set extracted from the knowledge graph and weighted by spatiotemporal correlation, which is different from the static rule set.

[0221] The necessary procedures are as follows:

[0222] Step S42-1: Construct an activated instance subgraph. Using the instance activated in step S41 as the seed node, perform K-hop Neighborhood Expansion in the knowledge graph. Typical values ​​are... Extract all nodes and edges within a 2-hop radius of the seed node to form an induced subgraph. .

[0223] Step S42-2: Perform graph neural network encoding on the induced subgraph. The same relational graph convolutional network architecture as in step S31 is used, with the input being... The output consists of node embeddings and edge embeddings. The classifier predicts the rule applicability based on the node embeddings: an applicability score. ,in The final embedding of the rule node (dimension 128). A multilayer perceptron for rule classification (64 hidden dimensions, ReLU activation function).

[0224] Step S42-3: Filter key semantic rules and generate a dynamic prediction rule set. Set an applicability threshold. Select The rule nodes constitute a dynamic prediction rule set. .right Each rule in the dataset is appended with its correlation strength, spatiotemporal correlation weight, and applicability score as metadata, forming an enriched rule object. For example, the applicability score of the rule "Severe damage to tunnel lining caused by strong earthquake in soft soil site" is... With a correlation strength of 0.80 and a spatiotemporal correlation weight of 0.75, it was included in the dynamic prediction rule set and given high priority.

[0225] Step S43: Start the multi-agent simulation model for parallel simulation. The structural agent, facility agent, environment agent, and personnel agent interact according to the dynamic prediction rule set. The environment agent synchronously inputs the ground motion parameters to drive the simulation timing.

[0226] The simulation timing refers to the time progression sequence of the multi-agent simulation model. In this application, it specifically refers to the set of discrete time steps based on the seismic motion input time axis, in which each agent advances synchronously according to a unified clock.

[0227] The necessary procedures are as follows:

[0228] Step S43-1: Initialize the simulation clock and global state. Set the simulation start time. (Unit: seconds), End Time (Unit: s), Time step s (unit: s). Global state vector ,in This is the vector of all structural damage indices. This is the vector of availability of all facilities. This is the coordinate matrix of all personnel positions (unit: m). This is the environmental ground motion acceleration vector (unit: m / s²).

[0229] Step S43-2, Environmental intelligent agent drives simulation timing. At each time step... Environmental intelligent agents read earthquake motion time history (Unit: m / s²), calculate the site response and broadcast it to the structural agent. The site response calculation adopts a one-dimensional equivalent linearization method: first, based on the site soil shear wave velocity... (Unit: m / s) and density (Unit: kg / m³) Calculate the shear modulus (Unit: Pa); then solve the wave equation to obtain the surface acceleration amplification factor. (Based on frequency) (Unit: Hz) variation); finally, the bedrock is input. Multiply Obtain the surface acceleration of the field (Unit: m / s²).

[0230] In step S43-3, each agent performs interactions based on the dynamic prediction rule set. The structural agent receives... Then, based on the structural damage rules matched in the dynamic prediction rule set (such as "if..."), (Approximately 1.96 m / s²), thus increasing the damage index growth rate by 30%). The damage evolution parameters are adjusted; the facility agent adjusts the state transition probability according to the facility degradation rules; the personnel agent adjusts the movement target and speed according to the evacuation rules. Each agent performs state updates in parallel within a unified time step and exchanges state information through a lightweight message queue.

[0231] Step S43-4: Advance the simulation clock and execute it cyclically. After each time step of global state update, the simulation clock advances. Repeat steps S43-2 and S43-3 until... Or it may trigger a global termination condition (such as the completion of the evacuation of all personnel or the collapse of a critical structure).

[0232] Step S44: Real-time acquisition of state change data generated by each agent during the simulation. Based on the state change data, semantic reasoning is performed through the knowledge graph to update the dynamic prediction rule set. Semantic reasoning is performed according to semantic relationships, so that the simulation state and prediction rules are updated synchronously.

[0233] Semantic reasoning refers to the process of logical deduction based on knowledge graph ontology and rules. In this application, it specifically refers to the computational process of using the current simulation state as facts, premise-conclusion form as rules, and executing forward or backward chain deduction through the reasoning engine.

[0234] The necessary procedures are as follows:

[0235] Step S44-1 involves detecting changes in the attribute values ​​of instances and the changes in the association strength of semantic relationships in the state change data, calculating the attribute value change rate and the association strength decay rate of the instances; when the attribute value change rate exceeds a preset mutation threshold, or the association strength decay rate is lower than a preset failure threshold, the corresponding instance is marked as an abnormal instance, and the corresponding semantic relationship is marked as a failed association. For the specific execution process of this step, please refer to step S441, which will not be elaborated here.

[0236] Step S44-2 involves extracting affected associated paths in the disaster chain based on the association between abnormal instances and failures using a knowledge graph, and calculating the path propagation loss of these associated paths. For a detailed explanation of the execution process of this step, please refer to step S442; it will not be elaborated upon here.

[0237] Step S44-3: Based on the path propagation loss, determine the degree of failure of the triggering conditions for each rule in the dynamic prediction rule set, and identify the rules to be evolved that have failed in their triggering conditions. For the specific execution process of this step, please refer to step S443; it will not be elaborated here.

[0238] Step S44-4 involves adjusting the trigger weights and execution priorities of the rules to be evolved based on the rules to be evolved, path propagation loss, attribute value change rate, and association strength decay rate, thereby generating an evolution prediction rule set. For a detailed explanation of the execution process of this step, please refer to step S444; it will not be elaborated upon here.

[0239] Step S44-5 involves synchronizing the evolutionary prediction rule set to the multi-agent simulation model, replacing the dynamic prediction rule set to drive the simulation evolution. For details on the specific execution process of this step, please refer to step S445; it will not be elaborated upon here.

[0240] Step S45: Based on the updated dynamic prediction rule set, aggregate the state change data output by each agent during the simulation process, calculate the service performance loss index, and output the quantitative evaluation results of structural damage and functional failure.

[0241] Among them, the service performance loss index refers to a numerical index that comprehensively reflects the degree of earthquake damage to underground transportation engineering. In this application, it specifically refers to a dimensionless scalar obtained by weighted aggregation of structural damage, facility failure and personnel casualty risk.

[0242] The necessary procedures are as follows:

[0243] Step S45-1: Aggregate the state change data of each agent. Extract the termination time from the simulation log. or the current moment Structural damage index vector Facility Function Availability Vector Personnel safety status vector (1 indicates safety, 0 indicates being trapped / injured).

[0244] Step S45-2: Calculate the sub-item loss indices. Structural damage loss indices. ,in For the first Weighting coefficients for each structural unit (determined by structural importance, e.g., 1.0 for the main tunnel and 0.8 for connecting passages); Facility function failure index ,in For the first Weighting coefficients for individual facility systems; personnel casualty risk indicators ,in For the first The weight coefficient of each personnel cluster.

[0245] Step S45-3: Calculate the overall service performance loss index. A weighted aggregation formula is used:

[0246] .

[0247] in, , , Global weight coefficients (typical values) The weights are consistent with the loss function weights in step S244-1, ensuring that the simulation evaluation and policy optimization objectives are aligned.

[0248] Step S45-4, output the quantitative assessment results. The assessment results are output as a structured report, including: (1) Structural damage level - based on Classification: Minor [0, 0.3), Moderate [0.3, 0.6), Severe [0.6, 0.9), Collapse [0.9, 1.0]; (2) Facility function failure level - based on Classification: Normal [0, 0.2), Degraded [0.2, 0.5), Partial Failure [0.5, 0.8), Complete Failure [0.8, 1.0]; (3) Comprehensive performance loss index (Values ​​[0,1]). For example, when a simulation terminates... (Moderate injury) (Downgrade) (Low risk) (Overall, moderate loss).

[0249] Real-time acquisition of state change data generated by each agent during simulation; based on the state change data, semantic reasoning is performed through a knowledge graph to update the dynamic prediction rule set, including:

[0250] Step S441: Detect changes in the attribute values ​​of instances and changes in the association strength of semantic relationships in the state change data, calculate the attribute value change rate of instances and the association strength decay rate of semantic relationships; when the attribute value change rate exceeds a preset mutation threshold, or the association strength decay rate is lower than a preset failure threshold, mark the corresponding instance as an abnormal instance and mark the corresponding semantic relationship as a failed association.

[0251] Among them, the attribute value change rate refers to the relative change in the instance attribute value per unit time, and in this application, it specifically refers to a high-order change indicator used to identify state mutations. The association strength decay rate refers to the rate at which the semantic relationship weight decreases per unit time, and in this application, it specifically refers to a negative change indicator used to identify the break in the disaster chain. The mutation threshold is the upper limit for determining whether an attribute value change constitutes an abnormal mutation. The failure threshold is the lower limit for determining whether the association strength has lost its effectiveness.

[0252] The necessary procedures are as follows:

[0253] Step S441-1: Calculate the rate of change of the instance's attribute values. For the structural damage index, the rate of change is calculated as follows: ,in The damage index at the current time step. The damage index at the previous time step. The simulation step size (unit: s).

[0254] The rate of change for facility availability is as follows:

[0255] .

[0256] Rate of change for personnel location coordinates (Characterizing relative displacement rate).

[0257] Step S441-2: Calculate the attenuation rate of semantic relation association strength. (Regarding association strength...) The attenuation rate is calculated as follows: (A positive value indicates attenuation). If Then define (Avoid division by zero).

[0258] Step S441-3: Perform anomaly detection and marking. Set the mutation threshold: structural damage index mutation threshold. s -1 Facility availability threshold s -1 Threshold for sudden changes in personnel location s -1 Set a failure threshold: correlation strength decay rate failure threshold. s -1 Decision logic: If or or If so, the corresponding instance is marked as an abnormal instance, and the abnormal flag is displayed. ;like Then the corresponding semantic relationship is marked as failure association, and the failure flag is... For example, if the damage index of a structural unit suddenly increases from 0.30 to 0.42 within 0.1 seconds, then... s -1 s -1 This structural unit is marked as an abnormal instance.

[0259] Step S442: Based on the association between abnormal instances and failures, extract the affected associated paths in the disaster chain through the knowledge graph, and calculate the path propagation loss of the associated paths.

[0260] In this context, the associated path refers to the directed edge sequence connecting disaster elements in a knowledge graph, and in this application, it specifically refers to the disaster propagation channel from the disaster-causing factor through the disaster-inducing environment and the disaster-bearing body to people. Path propagation loss refers to the cumulative degree of damage caused by the disaster propagating along the associated path, and in this application, it specifically refers to the dimensionless index of the state degradation of each node and the weakening of edge associations on the integrated path.

[0261] The necessary procedures are as follows:

[0262] Step S442-1: Using the anomaly instance as the seed node, perform a forward path search in the knowledge graph. Employ the Depth-First Search (DFS) algorithm, starting from the anomaly instance and traversing forward along the directed edges of semantic relationships. The search depth is limited to a certain value. (Characterizing the maximum propagation hops of a disaster chain), extract the set of all reachable paths. .

[0263] Step S442-2: Calculate the node state degradation loss for each path. Path node loss ,in, For nodes The weighting coefficients (determined by node type: disaster-causing factor 0.2, disaster-prone environment 0.2, disaster-bearing body 0.4, personnel 0.2) For nodes The state degradation index (damage index for disaster-bearing bodies, 1-safety state index for personnel, and normalized intensity index for disaster-causing factors and disaster-prone environment).

[0264] Step S442-3: Calculate the edge association weakening loss for each path. Edge loss ,in The current correlation strength, The side length coefficient (obtained by spatial distance normalization) , For nodes and Spatial distance (unit: m) The maximum spatial distance in the knowledge graph (unit: m)).

[0265] Step S442-4: Calculate the path propagation loss. (Comprehensive Loss) ,in, , Path loss weights (typical values) For example, the node loss of the path "Earthquake E1 → Soft Soil Site S1 → Tunnel T1 → Personnel H1". , edge loss Then path propagation loss .

[0266] Step S443: Based on the path propagation loss, determine the degree of failure of the triggering conditions of each rule in the dynamic prediction rule set, and identify the rules to be evolved that have failed in triggering conditions.

[0267] The trigger condition failure level refers to the degree to which the preconditions of a rule are no longer met due to changes in the disaster state. In this application, it specifically refers to a scalar that quantifies the decrease in the applicability of a rule, with path propagation loss as input. Rules to be evolved refer to rules whose trigger condition failure level exceeds a threshold and need to be eliminated or replaced from the rule base. In this application, it specifically refers to a subset of rules marked as failed in the dynamic prediction rule set.

[0268] The necessary procedures are as follows:

[0269] Step S443-1: Establish the mapping relationship between paths and rules. For the dynamic prediction rule set... Each rule in Extract the set of instances involved in its preconditions. Searching for information in a knowledge graph In the example, all associated paths of a node constitute the rules. Path Coverage Set .

[0270] Step S443-2: Calculate the failure level of the rule triggering condition. For the rule... Its degree of failure Calculated as a weighted average of path propagation loss:

[0271] .

[0272] in The number of paths in the path coverage set. The path weight is determined by the semantic relationship between the path and the rule; for example, a path directly referenced by the rule is weighted at 1.0, while an indirectly related path is weighted at 0.5. The value ranges from [0,1]. The larger the value, the more severe the failure of the rule triggering condition.

[0273] Step S443-3: Identify the rules to be evolved. Set the failure level threshold. .right All rules in the middle, if If the rule is not specified, it will be marked as a rule to be evolved and added to the set of rules to be evolved. And it will be downgraded from the current set of dynamic prediction rules (reduced in execution priority or suspended from triggering). For example, the path propagation loss weighted average of the rule "severe damage to tunnel lining caused by strong earthquake in soft soil site" This rule is marked as a rule to be evolved because the actual damage mode of the tunnel structure in the current simulation has deviated from the preset premise of this rule (such as the actual site being hard soil or the structure having been reinforced in advance).

[0274] Step S444: Based on the rules to be evolved, path propagation loss, attribute value change rate, and association strength decay rate, adjust the trigger weight and execution priority of the rules to be evolved to generate an evolution prediction rule set.

[0275] In this application, trigger weight refers to the probability or importance coefficient of a rule being triggered for execution, and specifically refers to a scalar used for rule ranking after dynamic adjustment. Execution priority refers to the order in which rules are executed when multiple rules compete, and specifically refers to integer levels or continuous scores in this application. The evolution prediction rule set refers to the new set of rules that has been adjusted to adapt to the current state of disaster evolution, and specifically refers to the updated version that replaces the original dynamic prediction rule set in this application.

[0276] The necessary procedures are as follows:

[0277] Step S444-1: Calculate the rule adjustment coefficients. (This applies to the rules to be evolved.) Comprehensive adjustment coefficient The calculation is as follows:

[0278] .

[0279] in, , , , Adjustment factor (typical value) ), To determine the degree of failure of the triggering condition, For rules The corresponding maximum path propagation loss, The normalized rate of change of the damage index. s -1 As a normalized benchmark; To normalize the correlation strength decay rate, s -1 This serves as the normalized benchmark. The larger the value, the more significant the adjustment to the rules is required.

[0280] Step S444-2, adjust the trigger weight. New trigger weight. ,in, The original trigger weight (determined by the spatiotemporal correlation weight in step S22). This is the weight decay coefficient (typically 0.8). If If the rule is not found, it will be removed from the rule set.

[0281] Step S444-3: Adjust execution priority. New priority ,in The original priority (as determined by the importance of the initial rule). This is the priority attenuation factor (typically 0.9). Priorities are arranged in descending order of value, with higher values ​​being executed first.

[0282] Step S444-4: Supplement with new rules to fill the gaps in the elimination rules. Based on candidate rules in the knowledge graph that are semantically similar to the removed rule, generate new rules through rule generalization: relax overly specific instance conditions in the original rule to conceptual conditions (e.g., relax "left line of tunnel in section XX" to "tunnel lining structure"), and assign initial trigger weights. With initial priority The adjusted rules are then merged with the new rules to generate an evolutionary prediction rule set. .

[0283] Step S445: Synchronize the evolution prediction rule set to the multi-agent simulation model and replace the dynamic prediction rule set to drive the simulation evolution.

[0284] Synchronization refers to the operation of pushing the evolutionary prediction rule set to the multi-agent simulation model and making it effective. In this application, it specifically refers to the atomic replacement mechanism of the rule set to ensure rule consistency during simulation. Replacement refers to the data operation of overwriting the old rule set with the new rule set. In this application, it specifically refers to the rule base update performed at the boundary of the simulation time step.

[0285] The necessary procedures are as follows:

[0286] Step S445-1: Perform rule set replacement at the simulation time step boundary. Set the rule update trigger time to... (Unit: s), where, It is an integer. This is the rule update cycle (typically 1.0s, meaning it updates every 10 simulation steps). At any given moment, pause the execution of actions by each agent and lock read / write permissions to the rule base.

[0287] Step S445-2: Perform an atomic replacement operation. Replace the old rule set in the rule cache of the multi-agent simulation model. Marked as expired, added to the new evolution prediction rule set. Update the rule index table and priority queue. After the replacement is complete, release the read-write lock and resume agent action execution.

[0288] Step S445-3: Verify the consistency of the replaced rules. Sampling checks are performed on the consistency between the preconditions and the current simulation state of five high-priority rules, requiring a consistency score. If consistency is insufficient, the system will roll back to the old rule set and trigger an alarm; if consistency is satisfied, a rule replacement log will be recorded, including the replacement time, changes in the number of rules, and a list of high-priority rules.

[0289] After generating the evolutionary prediction rule set, the method also includes a co-evolutionary step involving the pattern layer and the interaction topology:

[0290] Step S44A: parse the evolution prediction rule set and check whether the evolution prediction rule set contains cross-domain coupled disaster elements not defined in the ontology; cross-domain coupled disaster elements include secondary disaster nodes or cross-domain association relationships.

[0291] Among them, cross-domain coupled disaster elements refer to disaster nodes or relationships that exceed the current ontology definition and involve multiple disaster types or multiple engineering domain interactions. In this application, they specifically refer to secondary disaster nodes (such as fires, floods, and toxic gas leaks caused by earthquakes) or cross-domain association relationships (such as the coupling effect of tunnel structure damage and the collapse of surrounding buildings).

[0292] The necessary procedures are as follows:

[0293] Step S44A-1, for the evolution prediction rule set Each rule in the process undergoes an ontology compliance check. The entity concepts and relationship types in the rule's preconditions and conclusions are extracted and compared with the current ontology. The set of concepts in With relation set Perform a comparison. If all concepts in the rule... And all relationships If a concept exists, it is determined to be a domain-specific rule; or relationship If so, it is determined to contain cross-domain coupled disaster elements.

[0294] Step S44A-2 involves classifying and labeling the identified cross-domain coupled disaster elements. Secondary disaster nodes are labeled as Secondary_Hazard, with subtypes including: Fire, Flood, Gas Leak, and Landslide. Cross-domain relationships are labeled as Cross_Domain_Link, with subtypes including: Structure-Building, Underground-Surface, and Engineering-Environment. For example, the concept "flooding" in the rule "If the width of the tunnel lining crack is greater than 5mm, it will induce groundwater seepage leading to flooding" is not in the current concept set. In this context, the relationship "inducing" is not present in the current set of relationships. In this context, the rule is marked as containing cross-domain coupled disaster elements (secondary disaster node: flooding).

[0295] Step S44B: If cross-domain coupled disaster elements are included, add the corresponding secondary disaster node types and cross-domain association definitions to the model layer and update the ontology.

[0296] Here, "supplementation" refers to the operation of adding new concepts or relations to an existing ontology, and in this application, it specifically refers to the knowledge graph editing process that formally expands the coverage of the ontology. "Secondary disaster node type" refers to the conceptual classification of secondary disasters in the ontology, and in this application, it specifically refers to the definition of disaster types introduced as a subclass of "disaster-causing factor" or an independent conceptual class.

[0297] The necessary procedures are as follows:

[0298] Step S44B-1: Define the secondary disaster node type. In the model layer ontology, add the concept class "Secondary Disaster" as a subclass of "Causing Factor." Its subordinate concepts include: Fire (attributes: fire source location (unit: m), spread rate (unit: m / s), heat release rate (unit: kW)), Flooding (attributes: water level (unit: m), flow velocity (unit: m / s), flood range (unit: m²)), Toxic Gas Leakage (attributes: leak source location (unit: m), concentration (unit: ppm), diffusion coefficient (unit: m² / s)), Landslide (attributes: landslide volume (unit: m³), ​​sliding velocity (unit: m / s), impact range (unit: m²)).

[0299] Step S44B-2: Define cross-domain relationships. Add a new set of relationship types to the schema layer ontology. This includes: inducing (domain: disaster-bearing body, value range: secondary disaster), exacerbating (domain: secondary disaster, value range: disaster-bearing body), and coupling (domain: any disaster element, value range: any disaster element, representing a strong bidirectional correlation). Define attributes for each new relationship: correlation strength, influence distance (unit: m), and duration (unit: s).

[0300] Step S44B-3: Perform ontology version update and consistency verification. OWL ontology version control mechanism is used to generate a new version of the ontology. ,in For the newly added concept set, This is to add a new set of axioms (such as the axiom of the temporal relationship between secondary disasters and primary disasters). This is the updated hierarchical structure. A consistency check is performed using the ontology inference engine: it checks whether newly added concepts and relationships cause class hierarchy cycles, attribute domain scope conflicts, or axiom contradictions. If the check passes, the new version of the ontology is released and synchronized to the knowledge graph; if the check fails, the conflict definitions are corrected and the ontology is resubmitted.

[0301] Step S44C: Based on the updated ontology, reconstruct the interaction topology and interaction permissions between the agents; the interaction topology includes the newly added coordination links and conflict coordination boundaries between agents.

[0302] In this context, "interaction topology" refers to the network structure of communication links and cooperative relationships between agents in a multi-agent simulation model, specifically a directed graph with agents as nodes and message channels as edges. "Interaction permissions" refers to the access control rules for instruction transmission, state sharing, and negotiation decisions among agents, specifically a read / write permission matrix defined based on agent type and hierarchy in this application.

[0303] The necessary procedures are as follows:

[0304] Step S44C-1: Create corresponding agent types based on the newly added ontology concept. For the secondary disaster node "Fire", a new FireAgent is added, responsible for simulating fire source spread, heat radiation, and smoke diffusion; for "Flooding", a new FloodAgent is added, responsible for simulating water level rise, flow velocity field, and flooding range; for "Toxic Gas Leakage", a new GasAgent is added, responsible for simulating concentration field and diffusion range; for "Landslide", a new LandslideAgent is added, responsible for simulating landslide movement and impact.

[0305] Step S44C-2: Reconstruct the interaction topology. Based on the existing four types of agents (environment, facilities, structure, and personnel), a new secondary disaster agent is added to the interaction network: the environment agent broadcasts primary ground motion parameters and site conditions to the secondary disaster agent; the secondary disaster agent broadcasts secondary loads (such as fire heat flux density (kW / m²), flood hydrostatic pressure (Pa), and toxic gas concentration (ppm)) to the structure agent; the secondary disaster agent broadcasts the danger zone range and evacuation guidelines to the personnel agent; and coupling links are established between secondary disaster agents (e.g., interaction between the fire agent and the flood agent: high-temperature evaporation effect). The interaction topology uses an adjacency matrix. It indicates that the dimension is ,in The total number of agent types, matrix elements This indicates the existence of an intelligent agent. arrive Message channel This indicates that there is no direct communication.

[0306] Step S44C-3: Update the interaction permission matrix. Based on the disaster response command chain, set the permission rules: the environmental agent retains broadcast permissions to all agents (including the secondary disaster agent); the secondary disaster agent has the permission to issue warnings and load data to the structure, facility, and personnel agents, but does not have the permission to receive high-level dispatch instructions (secondary disasters are a passive evolution process); the structure, facility, and personnel agents have the permission to provide feedback on local conditions (such as structural temperature, facility operating status, and personnel location) to the secondary disaster agent for calibrating the secondary disaster evolution model.

[0307] Step S44D: Based on the reconstructed interaction topology, a fast adaptation strategy for each agent is generated through a preset meta-learning algorithm; the fast adaptation strategy includes interaction protocol update and local reward function redefinition.

[0308] Meta-learning algorithms refer to machine learning paradigms that learn how to learn. In this application, they specifically refer to algorithmic frameworks that enable agents to quickly adapt to new tasks or environments, such as Model-Agnostic Meta-Learning (MAML) or gradient-based meta-learning. Fast adaptation strategies refer to a set of strategies that allow agents to rapidly adjust their behavior patterns after sudden changes in interaction topology or environment. In this application, they specifically refer to a policy package that includes updates to interaction protocols and redefinition of local reward functions.

[0309] The necessary procedures are as follows:

[0310] Step S44D-1: Construct a meta-learning task set. Define each type of secondary disaster as a new task. Task set For each task, historical simulation trajectories are collected as a support set. With QuerySet Each set contains 100 trajectories (each trajectory is a state-action sequence of a complete simulation process).

[0311] Step S44D-2: Perform MAML meta-training. This involves using the original policy network parameters of each agent. Using the initial values, perform gradient updates in the inner loop on the support set: ,in This is the step size of the meta-learning inner loop (typically 0.01). For the task The loss function is applied (in the same form as the loss function in step S244). Then, an outer loop meta-gradient update is performed on the query set: ,in This is the outer loop step size for meta-learning (typically 0.001). After 10,000 rounds of meta-training, the pre-trained parameters are obtained. This parameter has the ability to quickly adapt to new secondary disaster missions.

[0312] Step S44D-3: Generate a rapid adaptation strategy package. When cross-domain coupled disaster elements are detected, each agent bases its strategy on... Perform a small number of gradient steps (e.g., 5 steps) to fine-tune and obtain task-specific parameters. The adaptation strategy package includes: (1) Interaction protocol update - based on the reconstructed interaction topology, update the message parser and state encoder to identify the new agent message type; (2) Local reward function redefinition - add secondary disaster-related terms to the original reward function, such as adding thermal damage penalty to the structural agent. (in This represents the actual heat flux density (unit: kW / m²). To allow for heat flux density (unit: kW / m²), human agents are penalized for exposure to toxic gases. (in This represents the actual concentration of toxic gas (unit: ppm). (This refers to the safe concentration limit (unit: ppm)).

[0313] Step S44E: Based on the rapid adaptation strategy, interaction permissions, and conflict coordination boundaries, drive the multi-agent simulation model to perform cross-domain coupled disaster simulation.

[0314] In this context, the conflict coordination boundary refers to the spatial or logical boundary by which multiple agents resolve target conflicts in a cross-domain coupled disaster scenario. In this application, it specifically refers to the decision authority boundary defined by the environmental agent based on the reconstructed interaction topology. Cross-domain coupled disaster simulation refers to the multi-agent simulation process that simultaneously simulates the primary disaster (earthquake) and secondary disasters (fire, flooding, etc.) and their coupling effects. In this application, it specifically refers to the extended simulation mode driven by the updated interaction topology and adaptation strategy.

[0315] The necessary procedures are as follows:

[0316] Step S44E-1: Load the quick adaptation policy package and the updated interaction topology. Each agent reads from the policy library. Update local policy network weights; read adjacency matrix from interactive topology library. With permission matrix Update the communication routing table and access control list.

[0317] Step S44E-2: Initialize the states of the secondary disaster agents. The fire agent initializes the fire source location and heat release rate based on the structural damage state (such as crack width and ventilation conditions); the flood agent initializes the seepage point location and flow rate based on geological and hydrological conditions and structural damage state; the toxic gas agent initializes the leak source and concentration field based on the facility damage state (such as pipeline breakage); and the landslide agent initializes the landslide parameters based on the seismic intensity and slope geological conditions.

[0318] Step S44E-3 executes the cross-domain coupled simulation loop. At each simulation time step, the primary disaster agents (environment, structure, facilities, personnel) proceed according to their original time sequence; the secondary disaster agents update their own evolutionary state based on the primary disaster state and broadcast secondary loads and hazardous areas to relevant agents; each agent re-evaluates the benefits of actions based on the local reward function in the rapid adaptation strategy, and negotiates and makes decisions through the updated interaction topology. For example, in... At time s, the structural agent detects that the crack width in a certain section reaches 5mm. Based on this, the fire agent initializes the fire source in this section, calculates the heat release rate of 500kW, broadcasts the heat flux density of 2.5kW / m² to the structural agent, and broadcasts the radius of the high-temperature danger zone of 10m to the personnel agent. The personnel agent adjusts the evacuation route to avoid the high-temperature zone according to the thermal damage penalty in the adaptation strategy. The structural agent calculates the comprehensive damage index by superimposing temperature stress and seismic stress according to the structural thermo-mechanical coupling model in the adaptation strategy.

[0319] Step S44E-4 involves continuous monitoring and dynamic adjustment of cross-domain coupling effects. The environmental agent, acting as a coordination hub, monitors the coupling strength between primary and secondary disasters: coupling strength index... ,in The weights for secondary disasters are (typical values ​​are 0.4 for fire, 0.3 for flooding, 0.2 for toxic gas, and 0.1 for landslide). For the first Secondary disaster intensity index (output by the secondary disaster agent). This represents the highest intensity of secondary disaster in history. If If this occurs, a global emergency response upgrade is triggered: the environmental agent forcibly increases the decision-making frequency of all agents (the time step is shortened from 0.1s to 0.05s), and backup computing resources are activated to ensure the real-time performance of the simulation.

[0320] Based on the same inventive concept, embodiments of the present invention provide a dynamic prediction system for earthquake damage in underground transportation engineering, including a memory and a processor, wherein the memory stores information that can be run on the processor to implement the following... Figure 1 The procedure for the method shown.

[0321] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for dynamically predicting seismic damage of underground traffic engineering, characterized in that, include: Based on data reflecting the characteristics of earthquake disasters in underground transportation projects, a panoramic knowledge graph of earthquake disasters in underground transportation projects is constructed. The knowledge graph includes a schema layer and a data layer. The schema layer adopts an ontology to describe the disaster-bearing body, disaster-causing factors, disaster-prone environment, personnel and their semantic relationships in the earthquake disaster environment. The data layer contains instance data corresponding to the ontology. The instance data includes instance attributes and semantic relationships between instances. The semantic relationships have an initial association strength defined according to the ontology. Based on knowledge graphs, a multi-agent simulation model for underground transportation engineering is established. The multi-agent simulation model includes a structural agent that simulates the mechanical response of the structure, a facility agent that simulates the functional state of the facility, an environmental agent that simulates the propagation of seismic motion, and a personnel agent that simulates the emergency behavior of personnel. Each agent interacts according to the rules provided by the knowledge graph. A bidirectional coupling mechanism is established between the knowledge graph and the multi-agent simulation model. The semantic rules reflecting the evolution logic of the disaster chain in the knowledge graph are converted into the initialization parameters and behavioral constraints of the multi-agent simulation model. The initialization parameters include structural mechanical parameters, facility status parameters, and personnel distribution parameters. At the same time, the state change data during the simulation process is collected in real time through a preset event-driven architecture and fed back to the knowledge graph to update the association strength of instance attributes and semantic relationships. After the bidirectional coupling mechanism is established, the seismic motion parameters are input based on the knowledge graph and the multi-agent simulation model. Semantic reasoning and parallel simulation of the multi-agent simulation model are performed through the knowledge graph to dynamically predict the service performance loss state of underground transportation engineering and output the quantitative assessment results of structural damage and functional failure.

2. The method of claim 1, wherein, The agents interact according to the rules provided by the knowledge graph, including: Semantic rules for earthquake disasters are extracted from the rules provided by the knowledge graph to obtain a set of rules. Semantic rules for earthquake disasters reflect the disaster-causing factors, disaster-inducing environment, disaster-bearing bodies and people. The disaster-causing relationships constitute the semantic basis of the disaster chain. The spatiotemporal correlation weights of each rule in the set of rules are calculated by a pre-defined spatiotemporal graph attention network. Weighted rules are obtained based on the spatiotemporal correlation weights, and key disaster chains with spatiotemporal correlations are identified based on the spatiotemporal correlation weights. The weighted rules are converted into a hierarchical behavioral strategy library, which is divided into three levels according to decision-making authority: high-level decision-making, mid-level allocation, and low-level control. The high-level decision-making strategy is executed by the environmental agent, used to coordinate earthquake propagation and disaster chain evolution; the mid-level allocation strategy is executed by the facility agent, used to manage emergency resource scheduling and facility function switching; and the low-level control strategy is executed by the structural agent and the personnel agent, respectively simulating structural damage response and personnel emergency actions. Structural damage response includes determining structural reinforcement areas, and personnel emergency actions include planning evacuation routes. The structural agent and the personnel agent execute a coordinated response based on the weighted rules. Based on a hierarchical behavior policy library, parallel interaction between agents is executed through hierarchical multi-agent reinforcement learning. Among them, the high-level policies of the environment agent drive the mid-level allocation of the facility agent, the mid-level allocation of the facility agent coordinates the low-level control of the structure agent and the personnel agent, and the state of each agent is synchronized in real time through a preset communication mechanism. During the parallel interaction process, a pre-set game theory negotiation mechanism is introduced to resolve the goal conflict between agents. The environmental agent acts as the coordination center. In response to the spatial competition between personnel evacuation routes and structural reinforcement areas, the equilibrium strategy combination is obtained by solving the Nash equilibrium. The environmental agent coordinates the conflict and optimizes the strategy combination based on the high-level decision-making strategy to generate a collaborative strategy.

3. The method for dynamic prediction of seismic damage in underground transportation engineering according to claim 2, characterized in that, Based on a hierarchical behavior policy library, parallel interactions between agents are executed through hierarchical multi-agent reinforcement learning, including: The environmental agent executes high-level decision-making strategies and issues mid-level deployment instructions to the facility agent. The facility agent coordinates the structural agent and the personnel agent to execute low-level control strategies according to the instructions, and respectively executes structural damage response and personnel emergency actions. The interaction permissions between the agents are determined based on the decision-making permissions corresponding to the three levels. Based on interactive permissions, a hierarchical reinforcement learning architecture corresponding to the hierarchical behavior policy library is constructed. This architecture includes: a high-level policy network for executing high-level decision-making strategies; a mid-level coordination network for executing mid-level allocation strategies; and a low-level execution network for executing low-level control strategies. The high-level policy network outputs a global disaster response objective, the mid-level coordination network breaks down the global disaster response objective into resource allocation sub-objectives, and the low-level execution network executes specific actions based on these sub-objectives. These actions include: the structural agent calculating quantitative indicators of structural damage response, including a structural damage index; and the personnel agent executing emergency actions, including planning evacuation routes. Each layer of the network makes collaborative decisions through a shared state space. The system transmits the agent status data generated by collaborative decision-making in real time through a preset lightweight message queue. The status data and agent actions are stored in the interaction log. The status data includes the structural damage index calculated by the structural agent, the facility resource reserve monitored by the facility agent, and the personnel location coordinates obtained by the personnel agent. With the optimization objective of minimizing the service performance loss of underground transportation engineering, including structural damage and functional failure, the network parameters of each layer are dynamically adjusted based on the agent's actions and state data recorded in the interaction log through a preset policy gradient algorithm.

4. The method of claim 3, wherein the method further comprises: The bidirectional coupling mechanism for establishing knowledge graphs and multi-agent simulation models includes: Using a pre-defined graph neural network algorithm, multi-dimensional features of semantic rules are extracted from the knowledge graph, and these features are converted into initialization parameters and behavioral constraints for a multi-agent simulation model. The graph neural network algorithm integrates an attention mechanism to prioritize the processing of disaster chains reflected by the semantic rules. Based on an event-driven architecture, the system collects state change data generated by each agent during the operation of the multi-agent simulation model in real time. The state change data includes structural damage index, facility resource reserves, and personnel location coordinates. The system processes the state change data through an incremental learning algorithm controlled by a time window, dynamically updates the attribute values ​​of instances corresponding to the ontology in the knowledge graph, and adjusts the association strength of semantic relationships between instances. The attribute values ​​and association strength are based on the ontology definition. Complete the bidirectional coupling between the knowledge graph and the multi-agent simulation model.

5. The method of claim 4, wherein the method further comprises: The system collects state change data in real time during the simulation process through a pre-defined event-driven architecture, feeds it back to the knowledge graph, and updates the association strength of instance attributes and semantic relationships, including: Define preset event triggering conditions, which include structural damage index exceeding preset damage threshold, personnel location coordinates deviating from preset baseline path by more than preset distance, and facility resource reserves falling below preset resource threshold; The system collects agent state data in real time according to event triggering conditions through a preset communication mechanism and filters it according to preset rules. The collected agent state data is divided into data blocks according to a preset time window for processing, and key features are extracted; the key features include deviation features from the preset baseline path. Using a pre-defined incremental learning algorithm, based on the processed data blocks, the attribute values ​​of instances corresponding to the ontology in the knowledge graph and the association strength of semantic relationships between instances are updated; the attribute values ​​and association strengths are based on the ontology definition. Synchronize the updated knowledge graph content to the multi-agent simulation model and verify the effectiveness of the update; The verified update results are associated with the state changes of the multi-agent simulation model and stored to form a closed-loop record.

6. The method of claim 5, wherein the method further comprises: Verifying the validity of the update includes: After synchronizing the updated knowledge graph content to the multi-agent simulation model, a preset single-step simulation is executed to obtain the prediction result at the current time step. The prediction result at the current time step is compared with the prediction result at the previous time step, and the prediction deviation is calculated. If the prediction deviation is less than or equal to the preset error threshold, the update is confirmed to be effective, and the updated knowledge graph content is used as the basis for the simulation of the next time step. If the prediction deviation exceeds the preset error threshold, the incremental learning algorithm's parameters are adjusted, and dynamic updates are re-executed until the error requirement is met or the preset number of iterations is reached.

7. The method of claim 1, wherein the method further comprises: Parallel simulation of semantic reasoning and multi-agent simulation models is performed using knowledge graphs to dynamically predict the service performance loss state of underground transportation engineering projects. The output quantitative assessment results of structural damage and functional failure include: Input seismic motion parameters, and based on the ontology in the knowledge graph, analyze the semantic relationships between seismic motion parameters and disaster-causing factors, disaster-prone environment, disaster-bearing bodies and people, and activate the instances corresponding to the semantic relationships. Based on the activated instances, key semantic rules are extracted using a pre-defined graph neural network algorithm to generate a dynamic prediction rule set; A multi-agent simulation model is launched for parallel simulation, in which the structural agent, facility agent, environment agent, and personnel agent interact according to the dynamic prediction rule set, and the environment agent synchronously inputs the ground motion parameters to drive the simulation time sequence. Real-time acquisition of state change data generated by each agent during the simulation; based on the state change data, semantic reasoning is performed through knowledge graph to update the dynamic prediction rule set; semantic reasoning is performed according to semantic relationships, so that the simulation state and prediction rules are updated synchronously. Based on the updated dynamic prediction rule set, the state change data output by each agent during the simulation process is aggregated to calculate the service performance loss index and output the quantitative assessment results of structural damage and functional failure.

8. The method of claim 7, wherein the method further comprises: Real-time acquisition of state change data generated by each agent during simulation; based on the state change data, semantic reasoning is performed through a knowledge graph to update the dynamic prediction rule set, including: The system detects changes in the attribute values ​​of instances and changes in the association strength of semantic relationships in the state change data, calculates the rate of change of attribute values ​​of instances and the rate of decay of association strength of semantic relationships; when the rate of change of attribute values ​​exceeds a preset mutation threshold, or the rate of decay of association strength is lower than a preset failure threshold, the corresponding instance is marked as an abnormal instance and the corresponding semantic relationship is marked as a failed association. Based on the association between abnormal instances and failures, the affected associated paths in the disaster chain are extracted through a knowledge graph, and the path propagation loss of the associated paths is calculated. Based on the path propagation loss, determine the degree of failure of the triggering conditions of each rule in the dynamic prediction rule set, and identify the rules to be evolved that have failed in their triggering conditions. Based on the rules to be evolved, path propagation loss, attribute value change rate, and association strength decay rate, the trigger weight and execution priority of the rules to be evolved are adjusted to generate an evolution prediction rule set; The evolution prediction rule set is synchronized to the multi-agent simulation model, replacing the dynamic prediction rule set that drives the simulation evolution.

9. A dynamic prediction system for earthquake damage in underground transportation engineering, characterized in that, It includes a memory, a processor, and a program stored in the memory and executable on the processor, which, when loaded and executed by the processor, implements the method for dynamic prediction of earthquake damage in underground transportation engineering as described in any one of claims 1 to 8.

Images