Disaster evolution path deduction method and system based on multi-modal space-time atlas

Abstract

The invention provides a disaster evolution path deduction method and system based on a multi-modal space-time atlas, and the method comprises the steps: obtaining the attribute data of a shopping mall, and constructing a deduction static atlas based on the attribute data of the shopping mall; the deduction static map initializes a state based on a set instruction, updates based on physical simulation, and outputs a deduction dynamic map; generating a comprehensive risk matrix based on the deduction static atlas and the deduction dynamic atlas, and clustering the comprehensive risk matrix to obtain a classification region cluster; performing physical prediction and digital prediction according to the classification region cluster and the deduction dynamic map, respectively obtaining physical determination and judgment data and digital determination and judgment data, and obtaining spreading probability data through the physical determination and judgment data and the digital determination and judgment data; according to the method, re-deduction is carried out through an intervention instruction, a deduction method capable of carrying out step-by-step deduction and backtracking deduction is provided, a deduction dynamic map is fused, and the deduction accuracy is improved in a physical and digital cooperation mode, so that making of a better emergency plan is assisted.

Description

Technical Field

[0001] This invention relates to the field of extrapolation technology, and more specifically, to a method and system for extrapolating disaster evolution paths based on multimodal spatiotemporal maps. Background Technology

[0002] In the field of fire safety and emergency management of large shopping mall buildings, predicting the evolution path of fire disasters is a prerequisite for formulating efficient emergency plans and realizing command decisions.

[0003] Traditional disaster simulation methods rely on static models or human experience for situation assessment, which has limitations. Static simulation models cannot integrate real-time environmental data during a fire, resulting in a disconnect between digital simulations and physical reality. This makes it difficult to respond to sudden changes in the fire situation, such as the instantaneous situational changes caused by the opening and closing of fire doors and the activation of sprinklers, leading to distorted simulation results. Consequently, the emergency response plans may be inadequate. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for extrapolating disaster evolution paths based on multimodal spatiotemporal maps, the method comprising:

[0005] Obtain mall attribute data and construct a deductive static map based on the mall attribute data;

[0006] The static graph is initialized based on the set command and updated based on the physical simulation, and the dynamic graph is output.

[0007] A comprehensive risk matrix is ​​generated based on the inferred static map and the inferred dynamic map. The comprehensive risk matrix is ​​then clustered to obtain the classification region clusters.

[0008] Physical and digital predictions are performed based on the classification region clusters and the inferred dynamic map. Physical and digital determination data are obtained respectively, and the spread probability data is obtained through the physical and digital determination data.

[0009] Obtain intervention instructions, inject the intervention instructions into the static graph of the inference, and then perform a re-inference.

[0010] Furthermore, acquire mall attribute data, and construct a deductive static map based on the mall attribute data, including:

[0011] Obtain the mall's BIM data and corresponding attribute data separately, and then obtain the mall's attribute data through the mall's BIM data and corresponding attribute data.

[0012] Extract entity nodes from the mall's attribute data;

[0013] Node edges are constructed based on the diverse relationships between entity nodes;

[0014] Each entity node is assigned a feature vector based on the mall's attribute data, and a feature matrix is ​​constructed based on the feature vectors.

[0015] A deductive static graph is constructed using entity nodes, node edges, and feature matrices.

[0016] Furthermore, the method also includes:

[0017] The entity nodes include at least spatial nodes, boundary nodes, and facility nodes;

[0018] The node edges include at least connected edges, adjacent edges, containing edges, control edges, and radiating edges;

[0019] The feature vectors include at least structural feature vectors, material feature vectors, topological feature vectors, and protection feature vectors;

[0020] The feature matrix is ​​represented as a set of all feature vectors.

[0021] Furthermore, the static graph is initialized based on the set command and updated simultaneously based on physical simulation, outputting the dynamic graph, including:

[0022] The system acquires and parses the set instructions, generates ignition point, fire source parameters and environmental condition data and injects them into the static simulation map, and initializes the static simulation map.

[0023] The physical simulation engine is invoked to perform a global simulation of the inferred static spectrum, and a physical solver is set up to obtain the virtual dataset for each time step through global simulation and physical solver.

[0024] At the end of each time step, the inferred static graph is updated in real time based on the virtual dataset, and the inferred dynamic graph is obtained based on the inferred static graph of each time step.

[0025] Furthermore, the physical simulation engine is invoked to perform a global simulation of the inferred static graph, and a physical solver is set up. The virtual dataset for each time step is obtained through the global simulation and the physical solver, including:

[0026] Set up the computational domain and control volume, assemble process sub-models based on the physics simulation engine, and perform simulations through the process sub-models;

[0027] The process sub-models include the fire source model, the flue gas model, and the heat transfer model;

[0028] The heat transfer model includes the heat conduction model, the convective heat transfer model, and the thermal radiation model.

[0029] A set of simulation equations is constructed based on the control volume and flue gas model, and the initial conditions and time steps are obtained. A physical solver is then constructed based on the set of simulation equations, initial conditions and time steps.

[0030] During the simulation of the process sub-model, the virtual dataset at each time step is solved cyclically by the physical solver and the process sub-model.

[0031] Furthermore, a comprehensive risk matrix is ​​generated based on the inferred static and dynamic maps. The comprehensive risk matrix is ​​then clustered to obtain classification region clusters, including:

[0032] Dynamic features are extracted from the dynamic graph of the deduction, and static features are extracted from the static graph of the deduction, so as to obtain dynamic feature vectors and static feature vectors respectively.

[0033] The comprehensive risk matrix is ​​obtained based on dynamic and static feature vectors, and the comprehensive risk matrix is ​​clustered using a clustering algorithm to obtain initial clusters;

[0034] Semantic annotation is performed on the initial clusters to obtain the classification region clusters.

[0035] Furthermore, physical and digital predictions are performed based on the classification region clusters and the inferred dynamic map, respectively obtaining physical and digital certainty judgment data. Spread probability data is then obtained from these physical and digital certainty judgment data, including:

[0036] Execution priority strategy based on classification region clusters;

[0037] The physical simulation engine is invoked to perform local simulation and prediction on the dynamic spectrum of the deduction, and physical determination data is obtained.

[0038] The T-GCN model is invoked to perform digital simulation and prediction of the dynamic spectrum, and digital determination data is obtained.

[0039] By fusing physical deterministic data with numerical deterministic data, the probability of propagation to be determined is obtained.

[0040] Furthermore, the physics simulation engine is invoked to perform local simulation predictions on the dynamic spectrum of the deduced model, obtaining physical determination data, including:

[0041] The fire source-target node pairs are obtained by simulating dynamic graphs.

[0042] The process sub-model assembled by the physics simulation engine is used to simulate and calculate the fire source-target node pair to obtain inference data;

[0043] Physically deterministic judgment data is obtained based on the dynamic graphs and inferred data.

[0044] Furthermore, the T-GCN model is invoked to perform digital simulation predictions on the projected dynamic spectrum, obtaining digital deterministic judgment data, including:

[0045] The T-GCN model is invoked, and tensor data is constructed based on the inferred dynamic graph.

[0046] Tensor data is input into the T-GCN model, and the T-GCN model performs operations including at least relation weighting, forward propagation, and probability output to obtain initial judgment data.

[0047] Post-process the initial judgment data to obtain numerical confirmation judgment data.

[0048] Furthermore, embodiments of the present invention also provide a disaster evolution path prediction system based on multimodal spatiotemporal maps, including:

[0049] The static construction module is used to build and simulate static maps based on mall attribute data.

[0050] The dynamic building module is used to initialize the static inference graph and output the dynamic inference graph through physical simulation;

[0051] The region classification module is used to perform clustering based on the inferred static map and the inferred dynamic map to obtain the classification region clusters;

[0052] The probability prediction module is used to perform physical and numerical predictions based on the classification region clusters and the inferred dynamic map to obtain the spread probability data;

[0053] The intervention replay module is used to inject intervention commands into the static graph and replay the graph.

[0054] In this embodiment, a static projection map is first constructed based on mall attribute data. Then, the static projection map is initialized based on set instructions and updated simultaneously based on physical simulation, outputting a dynamic projection map. Next, clustering is performed based on the static and dynamic projection maps to obtain classification region clusters. Subsequently, physical and digital predictions are performed based on the classification region clusters and the dynamic projection map, and propagation probability data is obtained through physical and digital determination data. Intervention instructions are obtained, injected into the static projection map, and re-projected. This method provides a projection method that allows for step-by-step projection and retrospective projection. During the projection process, the dynamic projection map is integrated, and a physical and digital collaborative approach is used, improving the accuracy of the projection. Furthermore, technical personnel can intuitively analyze the data during the projection process and perform comparative analysis through retrospective analysis. This method enables the development of more efficient emergency plans. Attached Figure Description

[0055] Figure 1This is a flowchart of the steps of the disaster evolution path inference method based on multimodal spatiotemporal map of the present invention.

[0056] Figure 2 This is a flowchart of step A400 in the disaster evolution path inference method based on multimodal spatiotemporal map of the present invention.

[0057] Figure 3 This is a schematic diagram of the disaster evolution path prediction system based on multimodal spatiotemporal maps of the present invention. Detailed Implementation

[0058] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0059] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0060] The following is in conjunction with the instruction manual appendix. Figure 1 To be continued Figure 3 The present invention will now be described in detail.

[0061] This embodiment provides a disaster evolution path prediction method based on multimodal spatiotemporal maps, including the following steps:

[0062] Step A100: Obtain shopping mall attribute data and construct a deductive static map based on the shopping mall attribute data.

[0063] In this embodiment, step A100 is performed based on steps A101 to A105.

[0064] Step A101: Obtain the mall's BIM data and corresponding attribute data respectively, and obtain the mall's attribute data through the mall's BIM data and corresponding attribute data.

[0065] Step A102: Extract the mall attribute data to obtain entity nodes.

[0066] Step A103: Construct node edges based on the multi-dimensional relationships between entity nodes.

[0067] Step A104: Assign a feature vector to each entity node using the mall attribute data, and construct a feature matrix based on the feature vector.

[0068] Step A105: Construct a deductive static graph using entity nodes, node edges, and feature matrices.

[0069] In step A101, the mall attribute data is constructed by combining the mall's BIM data with the corresponding attribute data. The mall's BIM data is represented as the mall's building information model. The mall's BIM data is usually an IFC format file, which contains three-dimensional geometry, spatial division, construction type and some attribute information. The corresponding attribute data is usually a text file, which contains incomplete or undefined detailed attributes in the mall's BIM data, such as material combustion performance, equipment parameters, fire compartment details and other information that are strongly related to the simulation.

[0070] The IFC parser is invoked to parse the mall's BIM data, identifying and extracting spatial data with clearly defined boundaries and functions, such as shops, corridors, and stairwells. Simultaneously, it acquires spatial volume, boundary geometry, and spatial containment relationships, including building component data such as walls, doors, windows, floors, and ventilation ducts, as well as geometric dimensions, spatial locations, and connection relationships. It also acquires facility data such as fire sprinklers, smoke detectors, fire shutters, and smoke exhaust vents, along with their installation locations and associated spaces. For attribute data, semantic parsing is performed after cleaning to identify and extract component IDs, equipment IDs, space IDs, component types, material data, equipment technical parameters, and space names. In practice, if the mall's BIM contains such attribute data, it can also be extracted from the mall's BIM. After extraction, the attribute data is matched and associated with the mall's BIM data, for example, matching space IDs with corresponding spatial data. Simultaneously, it ensures that all spatial coordinates are converted to the same global coordinate system and that the units of measurement are the same. The extracted and matched data constitutes the mall's attribute data.

[0071] In step A102, based on the needs of the shopping mall fire disaster simulation, entity nodes are defined. These are categorized into three types: spatial nodes (including shops, corridors, atriums, etc.), boundary nodes (including vertical partitions (e.g., walls), horizontal partitions (e.g., floors), connecting components (e.g., doors, windows), and pipe components (e.g., ventilation ducts), and facility nodes (e.g., fire-fighting facilities (e.g., fire sprinklers, fireproof roller shutters, smoke exhaust vents)). Once defined, all objects can be extracted from the shopping mall attribute data as spatial nodes, boundary nodes, and facility nodes, and a node pair can be created for each entity. Each node object is assigned a unique ID and its inherent attributes are assigned, such as geometric dimensions and material data. These attributes are then separated into static and dynamic attributes. Static attributes are those that do not change during a fire or are forcibly changed by external events, such as material data and geometric dimensions. Dynamic attributes are those that change during a fire or are forcibly changed by external events, such as trigger status, open / closed status, and current temperature. Then, spatial location attributes, such as spatial location, installation location, and connection relationships, can be assigned to each node. At this point, the physical node component is complete.

[0072] In step A103, each relationship type explicitly defines its logical meaning, applicable node types, directionality, and attributes. These can include connectivity, adjacency, containment, control, and radiation relationships. For connectivity, it means two spatial nodes are physically directly connected. Attributes include the connecting medium (e.g., door, fireproof roller shutter), effective flow area, and default open / closed state. Directionality is usually undirected, but can be considered directional depending on the state of the connecting medium. For adjacency, it means two spatial nodes are separated by a common partition, such as a wall, but can be connected through thermal conduction. Interactions, with attributes such as thermal conductivity, thickness, and partition components, are undirected. For containment relationships, they represent a subordinate or containment relationship, such as a space containing facilities, with the attribute being containment type, and the directionality is directional. For control relationships, they represent logical control, such as a smoke detector controlling a fireproof roller shutter, with the attribute being control logic, and the directionality is directional. For radiation relationships, they represent the possible paths through which a combustion point can affect another node via radiation, even in the case of non-direct contact, with attributes such as viewing angle coefficient and radiation attenuation factor, and the directionality is directional, from source to target.

[0073] Next, iterate through all door / window nodes. For each door / window node, using the two connected ID attributes, create a connecting edge between the two connected spatial nodes. The connecting medium attribute of the connecting edge points to the door / window node. The effective flow area is taken from the geometric dimensions of the door. Perform three-dimensional spatial position determination on all spatial nodes. If two spatial nodes share a partition component, create an adjacent edge between them. The partition component attribute of the adjacent edge points to the wall / floor node. The thermal conductivity and thickness are obtained from the static attributes of the component node. For each facility node, based on its ID, create an containing edge between the facility node and its associated spatial node. According to the fire protection design logic, connect the smoke detector node to the fire protection facility node it controls. A control edge is created between the nodes. The control logic of the control edge can be defined as triggering the device to act when the detector alarms. Traverse all spatial nodes. If there is no obvious obstruction between the center of node A and a certain surface of node B, a radiation edge is created in the direction of A→B. Its viewing angle coefficient depends on the size of the doorway and the relative position of the two spaces. The radiation attenuation factor can be set in the range of 0-1. After obtaining the edges, the edge weights also need to be set. The edge weight of the connected edge can be set as effective flow area / reference flow area * coefficient. The edge weight of the adjacent edge needs to be normalized by the thermal conductivity and thickness, and can be set as (thermal conductivity + thickness) / coefficient. The edge weight of the radiation edge can be set as viewing angle coefficient * radiation attenuation factor. The edge weight of the control edge can be set to 0 or 1.

[0074] In step A104, each spatial node is traversed, and structural feature vectors are extracted directly from the geometric attributes of the nodes. These structural feature vectors represent the physical geometric attributes of the space, such as volume and net height. Material feature vectors represent the risk of the structural materials, which can be obtained through aggregate calculations, such as a weighted average of the ignition points and structural areas of various structural materials within a space. Protection feature vectors represent the quantified fire resistance within the space, including fire extinguishing index, smoke extraction index, detection index, and fire resistance index. The fire extinguishing index can be calculated... The ratio of the sum of the sprinkler protection areas in a space to the space area is obtained. The smoke exhaust index is obtained by calculating the ratio of the designed smoke exhaust volume of the relevant smoke exhaust outlets in a space to the space volume. The detection index is obtained by calculating the ratio of the number of detectors in a space to the area. The fire protection index can identify the connecting medium of all connecting edges connected to a space. If it is a fire protection facility, a protection coefficient is assigned. For example, if the fire resistance time of a fire-resistant roller shutter is 4 hours, the protection coefficient can be set to 1.4. If it is 2 hours, it can be set to 1.2. Then the protection coefficients in the space can be summed to obtain the fire protection index.

[0075] In addition to the attribute feature analysis mentioned above, a topological feature vector is introduced. The topological feature vector represents the positional importance of a node in the global and local graph. It is obtained through graph theory algorithms. Before this, spatial nodes with safe exits need to be marked. The topological structure of nodes and edges is input into the graph theory algorithm to calculate the degree centrality, node centrality, and compact centrality of each spatial node. Taking the spatial node with a safe exit as the endpoint, the topological shortest path length from each spatial node to the nearest safe exit is calculated. Weighted distance is used to evaluate the compactness of the neighborhood of each spatial node to obtain the clustering coefficient. A high clustering coefficient may indicate complex connectivity in the local area. The topological feature vector is constructed based on degree centrality, node centrality, compact centrality, topological shortest path length, and clustering coefficient. It is different from attribute feature analysis. The topological feature vector reveals the vulnerability in the global structure. For example, a node located in the center and connecting multiple areas will have a huge impact on the disaster once it catches fire.

[0076] Missing values ​​in the structural feature vector, material feature vector, topological feature vector, and protection feature vector are handled and standardized, for example, by using median filling and Z-score standardization. The structural feature vector, material feature vector, topological feature vector, and protection feature vector are then encapsulated into feature matrices. Static graphs are derived through entity nodes, node edges, and feature matrix components.

[0077] Step A200: The static graph is initialized based on the set command and updated based on the physical simulation, and the dynamic graph is output.

[0078] Step A203: At the end of each time step, the inference static map is updated in real time based on the virtual dataset, and the inference dynamic map is obtained based on the inference static map of each time step.

[0079] In this embodiment, step A200 is performed based on steps A201 to A203.

[0080] Step A201: Obtain and parse the set instructions, generate ignition point, fire source parameters and environmental condition data and inject them into the simulation static map, and initialize the simulation static map;

[0081] Step A202: Call the physical simulation engine to perform a global simulation of the deduced static spectrum, and set up the physical solver to obtain the virtual dataset for each time step through global simulation and physical solver.

[0082] In this embodiment, step A202 is performed based on steps A2021 to A2023.

[0083] Step A2021: Set up the computational domain and control volume, assemble the process sub-model based on the physics simulation engine, and perform simulation through the process sub-model.

[0084] Step A2022: Construct a set of simulation equations based on the control volume and flue gas model, and obtain the initial conditions and time steps. Construct a physical solver based on the set of simulation equations, initial conditions and time steps.

[0085] Step A2023: During the process sub-model simulation, the virtual dataset for each time step is solved cyclically by the physical solver and the process sub-model.

[0086] For step A201, the system first receives setting instructions for inputting the virtual fire scenario via a graphical interface, form, or script. These instructions include selecting the specific spatial location of the fire, the fire growth type (e.g., t² fire), the initial heat release rate, distribution assumptions, combustion characteristics, and the building's special state at the start of the simulation. For example, it assumes that the fire door in area B3 is intentionally kept open. After receiving the setting instructions, the system analyzes them. For the graphically selected fire location, its coordinates are transformed to three-dimensional coordinates in the building's global coordinate system. For the location described in text, such as the clothing store in area A, the system uses spatial node names from the simulated static map. The system locates specific nodes to obtain ignition point data. Qualitative descriptions of user-selected t² fire, initial heat release rate, distribution assumptions, and combustion characteristics are mapped to corresponding parameters in the physics simulation engine, such as a growth coefficient of 0.1876 kW / s². Direct numerical inputs can be used directly to obtain fire source parameter data. The system parses the environmental settings in the text and extracts the target entity to be modified, such as the opening status of the B3 fire door mentioned above, thereby obtaining environmental condition data. It should be noted that t² fire in the above text is a mathematical model used in fire dynamics to describe the growth of fire.

[0087] In the static inference map, all affected nodes are located. Using the ignition point data, the spatial node closest to the coordinate is found and used as the starting point of the main fire source. Based on the environmental condition data, the corresponding node is found in the map. A preliminary analysis is performed to identify the list of adjacent nodes that the main fire source node will directly affect through connected edges and adjacent edges in the initial state. This provides boundaries for the local refinement of the subsequent physical simulation. The combustion state of the main fire source node is changed from unburned to burned. Fire source and risk level labels are added to the node type. At the same time, a sub-object is created and associated with the node. This sub-object must contain the fire growth type, such as t² fire, the current heat release rate (initial value), combustion characteristic parameters, such as effective heat of combustion and smoke production rate, and geometric assumptions, such as central fire or wall fire, to calculate the initial flame shape and radiation. All of these can be obtained from the fire source parameter data. Then, the initial environmental state of the node, such as initial temperature and initial smoke, is calculated and set using the fire source parameter data.

[0088] Once the injection is complete, you can traverse all the nodes that need to be modified, change the state of the relevant nodes, and overwrite their default values. For example, change the current closed state of a fire door to the open state, and at the same time calculate the edge weights of the relevant edges.

[0089] In step A2021, the updated static map of the simulation is analyzed to determine the simulation range of the physical simulation. Usually, all spatial nodes of the entire building are used as the computational domain. However, for super-large buildings, a Region of Interest (ROI) can be dynamically set according to the location of the fire source and the ventilation path. This region is the focus of the calculation, while the region outside the ROI can use a simplified model to balance the requirements of accuracy and speed. Each spatial node is defined as a control volume, and its geometric parameters such as volume and area can be obtained directly from the node attributes. The edges between spatial nodes are transformed into exchange relationships between control volumes. Connecting edges are the flow openings between control volumes, with parameters including opening area, center height, and flow coefficient. Adjacent edges are the heat-conducting walls between control volumes, with parameters including common wall area, material thermal conductivity, and thickness. Radiation edges are the radiation viewing angle coefficients between control volumes, with parameters including viewing angle coefficient and attenuation factor. In some possible embodiments, secondary meshing can be performed inside the fire source node and its adjacent high-risk nodes to calculate the structure of the flame and hot smoke layer more finely and improve the simulation accuracy of key areas.

[0090] Once the computational domain and control volume are determined, appropriate physical process sub-models can be selected and assembled from the physical simulation engine (model library) according to the type of fire disaster and node status. Typically, this involves assembling and splicing fire source models, smoke models, and heat transfer models. The fire source model can be divided into a growth phase and a steady-state phase. The growth phase uses a t² fire model, with the heat release rate formula Q=a(t-t0)², where a is the fire growth coefficient, which can be selected according to the set rapid fire, medium-speed fire, etc., corresponding standard values, t0 represents the start time, and t represents the current time. When the fire grows to a certain extent and reaches the sprinkler suppression level, it switches to a steady-state model to maintain the heat release rate at a steady-state value. The fire source model is used to update the temperature and heat release rate of the fire source nodes. The smoke model uses a CFAST-type dual-zone model, where CFAST represents fire simulation software. Its modeling concept is to divide the building space into a dual-zone model with an upper hot smoke layer and a lower cold air layer. The CFAST-type dual-zone model here uses this modeling method. Within each spatial node, the smoke... A hot flue gas layer forms at the top, and a cold flue gas layer forms at the bottom, with a clear interface between the two layers. The flue gas model uses discrete equations from the subsequent physics solver and real-time state parameters from other models, such as wall convection and radiative heat flux, to solve and update the layer temperature, concentration, and interface height. The heat transfer model includes heat conduction, convective heat transfer, and thermal radiation models. The heat conduction model uses a one-dimensional steady-state heat conduction approximation, expressed as q1=U*A(T1-T2), where U represents the overall heat transfer coefficient, which can be obtained from the component nodes of adjacent edges. The thermal conductivity and thickness of the material are calculated, where A represents the area, and T1 and T2 are the nodal temperatures on both sides, respectively. The convective heat transfer model includes wall convection and open convection. Wall convection can be expressed using the empirical formula q2=h*A(T1-T2), where h is the convective heat transfer coefficient, usually taken as a constant or a simple function. Open convection is already included in the flue gas model. The thermal radiation model uses a simplified radiation model based on the viewing angle coefficient and the Stefan-Boltzmann law, with the radiative heat flux being q3=σ*F*d(T1). 4 -T2 4 ), where F is the viewing angle coefficient, d is the emissivity, which can be obtained from the material properties, and σ is a coefficient. The heat transfer model is used to update the heat transfer of the component nodes. At this point, the physical process sub-model splicing is completed.

[0091] For steps A2022 and A2023, the volume of the control volume is divided into an upper flue gas layer and a lower air layer, and discrete equations for mass conservation, energy conservation, and component conservation are constructed respectively. For the mass conservation discrete equations, mass balances are established for the upper flue gas layer and the lower air layer of each control volume. The mass change in the upper layer is equal to the sum of the mass of flue gas flowing in through the connecting opening, the lower air entrained by the fire plume, and the gas produced by fuel combustion. The mass change in the lower layer mainly depends on the opening flow and entrainment losses. The opening flow rate is determined by the geometry of the connecting edges. The properties and interlayer pressure difference calculated based on Bernoulli's principle determine the internal energy balance for the energy conservation discrete equation. The upper layer's energy changes include open-flow, all heat power released by the ignition source, received radiant heat, and wall convection. The lower layer mainly handles open-flow energy and wall cooling. Radiant heat calculations utilize the radiation edge viewing angle coefficient and surface emissivity. For the composition conservation discrete equation, the CO mass fraction is tracked using the same upper and lower layer architecture. Its changes are jointly determined by open-flow convection, entrainment, and ignition source generation rate. Ultimately, all control volumes are determined through its... The shared edge coupling is assembled into a global micro-discrete equation system to solve the process sub-model. The initial state of all control volumes is obtained by deriving the static graph. The dynamic state attributes of the nodes in the graph at the initial time are directly read and used as initial conditions. In some possible embodiments, the external environment of the building, such as outdoor temperature and wind force, can be set as fixed boundary conditions. Parameters related to the ventilation system can be added to the equations as commutative terms. Then, the time step is set. A physical solver is constructed by simulating the equation system, initial conditions, and time steps. Based on the state of each time step, the solver calculates the state variables of the control volume between two time steps due to the release of the fire source, fluid flow, heat and mass transfer. For example, based on the state of time step T0, the physical solver calculates the state variables of the control volume from time step T0 to time step T1 and extracts physical quantities such as average temperature. At the same time, data that can be directly obtained from the process sub-model, such as heat release rate, is also extracted. This is usually an iterative process until the solution of the equation system converges in the next time step.

[0092] Under the control framework of iterative solution, the solution is executed from one time step to the next. The physical solver solves the discretized mass, energy, and component conservation equations based on all state variables of the previous time step, and obtains the new state variable values ​​of each control volume in the next time step. In the next time step, the real-time status of the facility nodes is checked, such as the temperature at the location of each fire sprinkler. If the temperature reaches its activation temperature, the attribute of the fire sprinkler is changed to active, and its water output and coverage area are obtained. These need to be considered in subsequent solution processes. The calculated status of all spaces and facilities is encapsulated into a standardized data stream, i.e., a virtual dataset, according to a predefined format and frequency. For example, it may contain data describing the shopping mall fire process, such as average temperature, upper smoke temperature, sprinkler activation, and maximum temperature.

[0093] In step A203, the deduced static graph receives the virtual dataset from step A2023 in real time, updates the attributes of its own nodes, and recalculates the edge weights of the nodes, thereby achieving a triggered update of the deduced static graph and obtaining the deduced dynamic graph.

[0094] Step A300: Generate a comprehensive risk matrix based on the simulated static map and the simulated dynamic map, and cluster the comprehensive risk matrix to obtain the classification region clusters.

[0095] In this embodiment, step A300 is performed based on steps A301 to A303.

[0096] Step A301: Perform dynamic feature extraction on the dynamic graph and static feature extraction on the static graph to obtain dynamic feature vectors and static feature vectors respectively.

[0097] Step A302: Obtain the comprehensive risk matrix based on the dynamic and static feature vectors, and perform clustering operations on the comprehensive risk matrix using a clustering algorithm to obtain the initial clusters.

[0098] Step A303: Semantically label the initial clusters to obtain the classification region clusters.

[0099] In step A301, feature extraction is performed on the nodes in the dynamic simulation graph. A set of dynamic risk features is extracted for each spatial node to describe its real-time behavior during the fire disaster. These features can characterize the current state and change trend of the node and are divided into state features, evolution features, exposure features, and topological influence features. State features represent the instantaneous values ​​at the current moment, such as the current temperature and the burning state of the node. Evolution features represent the rate of change, such as the rate of temperature change. Exposure features represent external threats, such as the intensity of received thermal radiation. Topological influence features represent the dynamic role of the node in the structure, such as the current connectivity state. Feature extraction is also performed on the nodes in the static simulation graph. A set of static features, i.e., static attributes, is extracted for each spatial node to describe the static features that will not change or will be forcibly changed by external events during the fire.

[0100] In step A302, the static and dynamic feature vectors extracted in step A301 are fused to generate a comprehensive risk vector for each spatial node. Then, the comprehensive risk vectors of all spatial nodes are integrated into a comprehensive risk matrix. Next, a density clustering algorithm is applied to cluster the comprehensive risk matrix. During the clustering process, the algorithm automatically calculates the distance or similarity of nodes in the feature space, thereby dividing all spatial nodes into several clusters with similar internal features and different component features, i.e., initial clusters. This aggregates discrete point information into regional probabilities with clear commonalities, simplifying the cognitive dimension.

[0101] In step A303, the initial clusters output by the clustering algorithm are semantically labeled, assigning them semantic labels with clear fire disaster response significance. These labels can be based on manual pre-definition or automated analysis, including core combustion zone, smoke spread zone, high-threat zone, and relatively safe zone. The core combustion zone consists of open flame nodes, characterized by high temperature and high temperature rise rate. The smoke spread zone consists of nodes such as corridors and atriums, characterized by high smoke concentration, stable temperature changes, and connection of multiple areas. The high-threat zone consists of nodes that have not yet caught fire, but whose static and dynamic characteristics indicate that they may be ignited in a short time, such as rapid temperature rise and strong heat radiation. The relatively safe zone indicates that it is currently less directly affected by the fire, but may face risks due to the development of the situation. After semantic labeling of the initial clusters, classification region clusters are formed. In some possible embodiments, they can also be labeled in the shopping mall BIM model with different colors or patterns for easy and intuitive observation.

[0102] Step A400: Perform physical and digital predictions based on the classification region clusters and the inferred dynamic map, and obtain physical and digital determination data respectively. Then, obtain the spread probability data through the physical and digital determination data.

[0103] In this embodiment, step A400 is performed based on steps A401 to A404.

[0104] Step A401: Execute a priority strategy based on the classification region cluster.

[0105] Step A402: Call the physics simulation engine to perform local simulation prediction on the dynamic spectrum of the deduction and obtain physical determination data.

[0106] Step A403: Call the T-GCN model to perform digital simulation prediction on the inferred dynamic spectrum and obtain digital determination data.

[0107] Step A404: Merge the physical determination data and the digital determination data to obtain the probability data of the spread.

[0108] In this embodiment, step A402 is performed based on steps A4021 to A4023.

[0109] Step A4021: Obtain the fire source-target node pair by deriving the dynamic graph.

[0110] Step A4022: Simulate the fire source-target node pair using the process sub-model assembled by the physics simulation engine to obtain inference data.

[0111] Step A4023: Obtain physical determination data based on the deduced dynamic spectrum and inferred data.

[0112] In this embodiment, step A403 is performed based on steps A4031 to A4033.

[0113] Step A4031: Call the T-GCN model and construct tensor data based on the inferred dynamic graph.

[0114] Step A4032: Input the tensor data into the T-GCN model, and perform operations including relation weighting, forward propagation and probability output through the T-GCN model to obtain initial judgment data;

[0115] Step A4033: Post-process the initial judgment data to obtain numerical confirmation judgment data.

[0116] In step A401, the impact of the classified region clusters on the global situation is evaluated according to predefined rules. For example, the core combustion zone, as the source of the fire disaster, directly affects all downstream areas and can be given the highest priority. The smoke spread zone may accelerate the spread of the fire disaster and can be given a high priority. The high-threat zone is a key point that may cause drastic changes in the fire disaster and can also be given a high priority. The relatively safe zone can be given a low or medium priority. For the highest and high priority nodes, high-fidelity calculation is maintained when the process sub-model and the physical solver perform calculation simulation. For medium and low priority nodes, standard-fidelity calculation is used. Simplified sub-models can be selected from the physical simulation engine to assemble the process sub-model, thereby reducing the simulation calculation time and data volume. During the iterative execution of steps A200 to A400, the process sub-model and the physical solver only use high-fidelity simulation calculation from time step T0 to time step T1. In subsequent loops, the calculation is based on the priority strategy.

[0117] In step A4021, the dynamic graph and classification region clusters are queried to identify all nodes in the "burned" state. For each "burned" node, the edge relationships of the dynamic graph are used to identify nodes with first-level and second-level adjacencies in the "unburned" state, thereby generating a potential target node set. The target node set is then paired with the corresponding "burned" nodes to form a fire source-target node pair. This process requires consideration of priority and selective inclusion of second-level adjacency nodes.

[0118] In step A4022, for each fire source-target node pair, the edges between all fire source-target node pairs are identified. Using the process sub-model and the physical solver, the radiative heat flux and convective heat flux are calculated and obtained. The radiative heat flux and convective heat flux are vector- or algebraically superimposed. If there are multiple edges, they are summed to obtain the total theoretical heat flux received by the target node surface, i.e., the inferred data.

[0119] In step A4023, the inferred data is applied to the material properties of the target node to determine ignition. Material parameters can be read from the static properties of the target node, and then a simplified thermal response model for thin or thick materials can be used. For example, for solid combustibles, the time required for the surface temperature to rise to the ignition point can be estimated using a formula. This time is inversely proportional to the received heat flux and the material's thermal inertia. A commonly used calculation formula is:

[0120] ;

[0121] in, This represents the time required for the surface temperature to rise to the ignition point. Indicated as the ignition point of the material. This is expressed as the received heat flux. The current temperature of the target node is then compared with the estimated ignition time. Based on the current simulation time and the next time step, if the ignition time is less than the end time of the next time step, the target node will be ignited at the ignition time; otherwise, it will be ignited. A confidence interval is added for each judgment. For example, the ignition time can be used as the confidence interval. The confidence interval can be obtained by perturbing the input parameters of the ignition time, such as ±10% of the heat flux, which can be represented in the form of [8 minutes, 10 minutes], describing the estimated ignition time within the range of 8 to 10 minutes. Finally, an ignition judgment list is generated, which includes the nodes to be ignited in the future time step and the confidence interval, i.e., physically deterministic judgment data.

[0122] In step A4031, the T-GCN model is trained using historical shopping mall fire data to obtain a neural network model capable of processing spatiotemporal graph data. This model encodes the fire spread pattern. To adapt to the input requirements of the T-GCN model, the dynamic graph needs to be preprocessed. The state features of each node in the dynamic graph, such as temperature, and edge weights are normalized according to their respective historical mean and standard deviation to ensure the stability of the T-GCN model's input. Then, a fixed-length and continuous time truncation window is set, which can be set to the length of the time step. Its length must be consistent with the length used during model training. The dynamic graph within the window is then converted into a node feature matrix, consisting of n nodes * d-dimensional features and an adjacency matrix of n * n. By stacking the matrices over time, a three-dimensional spatiotemporal graph tensor is formed, i.e., tensor data, which serves as the direct input to the model.

[0123] T-GCN is a supervised learning model that requires input data to have a fixed format and statistical distribution. It eliminates dimensions through normalization to ensure that the model is not dominated by features with large numerical ranges. The sliding window provides the time dependency necessary for the model to perform analysis, because the spread of fire disaster is a dynamic process and its current state strongly depends on historical evolution. Tensorization is a prerequisite for the model's input.

[0124] In step A4032, the influence weights between nodes are dynamically adjusted based on the current node state and edge attributes to simulate real attention allocation. For each edge, an attention coefficient is calculated based on the characteristics of the source and target nodes and the edge attributes. For example, when node 1 has an extremely high temperature and node 2 is made of flammable material, the coefficient is larger, and vice versa. If there is a firewall between them, the coefficient is smaller. This can be accomplished by the attention mechanism. The edge weights in the deduced dynamic graph are weighted by the attention coefficient to form a dynamic weighted adjacency matrix. The dynamic weighted adjacency matrix reflects the connection relationship and the relationship of which influences which other node is more critical in the current state.

[0125] Tensor data and a dynamically weighted adjacency matrix are input into the T-GCN model, and a forward propagation is performed. Within each time window, the model aggregates the neighbor information of each node. For example, for node 1, its new representation is composed of a weighted combination of its own features and all its neighbors, with the weights coming from the dynamically weighted adjacency matrix. This can capture the spatial pattern of fire spreading from the surrounding area. At the same time, the model uses one-dimensional convolutions or recurrent units to slide along the time axis to analyze the changing trend of each node's features over time. This can capture the temporal pattern of the continuous rise in temperature leading to eventual ignition. After multiple layers of spatiotemporal convolution, the model's last layer classifier, such as a fully connected layer + Softmax, outputs the state transition probability and confidence interval of each node at a set future time point, i.e., the initial judgment data. For example, the probability of node A being ignited within the next 3 minutes is 0.78, and the ignition time is estimated to be within the range of 3 to 6 minutes.

[0126] In step A4033, after the initial judgment data is acquired, post-processing rules are applied. For example, a node that has been determined to be on fire needs to have its ignition probability forcibly set to 1, while the probability of a node that has been determined by physical simulation to be impossible to ignite under the current conditions should be limited to a very low upper limit to ensure that the digital prediction does not conflict with known facts or physical laws. This outputs an ignition judgment list, which includes the ignition probability of nodes in future time steps, i.e., digital determination judgment data. In some possible embodiments, the individual probabilities of nodes in a region can also be aggregated according to the region classification to generate the overall probability of each classification region, which can be used for data analysis of regional fire probability.

[0127] In step A404, after obtaining the physical and digital determination data, the physical and digital determination data are merged. For example, for a node, if the physical determination data determines that it will ignite, then regardless of the probability of the digital determination data, the final probability of it being ignited is 100%. If the physical determination data determines that it will not be ignited, then the final probability is determined by the probability of the digital determination data. Finally, a propagation probability data is generated for each edge, including the ignition probability and the expected propagation time interval. The expected propagation time interval can be obtained through the confidence interval.

[0128] Step A500: Obtain the intervention command, inject the intervention command into the static inference graph, and perform a re-inference.

[0129] In step A500, intervention instructions are manually set. For example, at a certain time step, all fireproof roller shutters between zones C and D are closed. At this time, the intervention action, intervention trigger time, and intervention target are extracted from the intervention instructions. As mentioned above, the intervention action is to close the fireproof roller shutters, the intervention trigger time is a certain time step, and the intervention target is all fireproof roller shutters between zones C and D. Subsequently, the dynamic graph is rolled back to the specified intervention trigger time to obtain the complete graph at this time. Based on the intervention instructions, the state and edge attributes of relevant nodes are modified. For example, the state attribute of the target fireproof roller shutter is changed from open to closed, and the edge weight of the control edge is set to 0, thereby obtaining the dynamic graph after intervention. Using this graph as the starting point, the process from steps A200 to A400 is re-executed to output the data after intervention. By comparing the data before and after intervention, it can be intuitively shown how the intervention changed the overall fire disaster pattern, which can be used to better analyze fire disasters and thus better formulate efficient emergency plans.

[0130] In this embodiment, a disaster evolution path prediction system based on multimodal spatiotemporal maps is also proposed, including:

[0131] The static construction module is used to build and infer static maps based on mall attribute data.

[0132] The dynamic construction module is used to initialize the static inference graph and output the dynamic inference graph through physical simulation.

[0133] The region classification module is used to perform clustering based on the inferred static map and the inferred dynamic map to obtain the classification region clusters.

[0134] The probability prediction module is used to perform physical and digital predictions based on the classification region clusters and the inferred dynamic map to obtain the spread probability data.

[0135] The intervention replay module is used to inject intervention commands into the static graph and replay the graph.

[0136] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A disaster evolution path deduction method based on a multi-modal spatiotemporal graph, characterized in that, The method comprises: acquiring mall attribute data, constructing a deduction static graph based on the mall attribute data; the deduction static graph is initialized based on a set instruction, and is updated based on physical simulation, and a deduction dynamic graph is output; a comprehensive risk matrix is generated based on the deduction static graph and the deduction dynamic graph, the comprehensive risk matrix is clustered, and a classified region cluster is acquired; physical prediction and digital prediction are performed based on the classified region cluster and the deduction dynamic graph, physical determination judgment data and digital determination judgment data are respectively acquired, and spread probability data is acquired through the physical determination judgment data and the digital determination judgment data; an intervention instruction is acquired, the intervention instruction is injected into the deduction static graph, and re-deduction is performed.

2. The multi-modal spatio-temporal graph-based disaster evolution path inference method according to claim 1, characterized in that, acquiring mall attribute data, constructing a deduction static graph based on the mall attribute data, comprising: respectively acquiring mall BIM data and corresponding attribute data, acquiring mall attribute data through the mall BIM data and the corresponding attribute data; extracting the mall attribute data to acquire entity nodes; constructing node edges based on the multi-element relationship between the entity nodes; assigning a feature vector to each entity node through the mall attribute data, and constructing a feature matrix based on the feature vector; constructing the deduction static graph through the entity nodes, the node edges and the feature matrix.

3. The multi-modal spatio-temporal graph-based disaster evolution path inference method according to claim 2, characterized in that, The method further comprises: wherein the entity nodes at least include space nodes, boundary nodes and facility nodes; wherein the node edges at least include connected edges, adjacent edges, containing edges, control edges and radiation edges; wherein the feature vectors at least include structure feature vectors, material feature vectors, topology feature vectors and protection feature vectors; wherein the feature matrix is represented as a set of all feature vectors. The deduction static graph is initialized based on a set instruction, and is updated based on physical simulation, and a deduction dynamic graph is output, comprising:

4. The multi-modal spatio-temporal graph-based disaster evolution path inference method according to claim 1, characterized in that, acquiring a set instruction and parsing, generating a fire starting point, fire source parameters and environmental condition data and injecting them into the deduction static graph, and initializing the deduction static graph; calling a physical simulation engine to globally simulate the deduction static graph, and setting a physical solver, acquiring a virtual data set of each time step through global simulation and the physical solver; at the end of each time step, the deduction static graph is updated in real time based on the virtual data set, and the deduction dynamic graph is acquired based on the deduction static graph of each time step. calling a physical simulation engine to globally simulate the deduction static graph, and setting a physical solver, acquiring a virtual data set of each time step through global simulation and the physical solver, comprising:

5. The method of claim 4, wherein, setting a calculation domain and a control body, assembling process sub-models based on the physical simulation engine, and simulating through the process sub-models; wherein the process sub-models include a fire source model, a smoke model and a heat transfer model; wherein the heat transfer model includes a heat conduction model, a convection heat transfer model and a heat radiation model; constructing a simulation equation set according to the control body and the smoke model, simultaneously acquiring initial conditions and time steps, and constructing a physical solver based on the simulation equation set, the initial conditions and the time steps; in the process of simulating the process sub-models, the physical solver and the process sub-models are used to cyclically solve the virtual data set of each time step. ​ 6. The multi-modal spatio-temporal graph-based disaster evolution path inference method according to claim 1, characterized in that, The comprehensive risk matrix is generated based on the deduced static graph and the deduced dynamic graph, the comprehensive risk matrix is clustered, and a classified region cluster is obtained, including: Dynamic feature extraction is performed on the deduced dynamic graph, and static feature extraction is performed on the deduced static graph, so as to obtain a dynamic feature vector and a static feature vector respectively; Based on the dynamic feature vector and the static feature vector, a comprehensive risk matrix is obtained, and clustering operation is performed on the comprehensive risk matrix by a clustering algorithm to obtain an initial cluster; The initial cluster is semantically labeled to obtain a classified region cluster.

7. The multi-modal spatio-temporal graph-based disaster evolution path inference method according to claim 1, characterized in that, Physical prediction and digital prediction are performed based on the classified region cluster and the deduced dynamic graph to obtain physical determination judgment data and digital determination judgment data respectively, and spread probability data is obtained based on the physical determination judgment data and the digital determination judgment data, including: Based on the classified region cluster, a priority strategy is executed; A physical simulation engine is called to perform local simulation prediction on the deduced dynamic graph to obtain physical determination judgment data; A T-GCN model is called to perform digital simulation prediction on the deduced dynamic graph to obtain digital determination judgment data; The physical determination judgment data and the digital determination judgment data are fused to obtain to-be-spread probability data.

8. The method of claim 7, wherein, A physical simulation engine is called to perform local simulation prediction on the deduced dynamic graph to obtain physical determination judgment data, including: A fire-source-target node pair is obtained through the deduced dynamic graph; A process sub-model assembled by the physical simulation engine is used to simulate and calculate the fire-source-target node pair to obtain inference data; Based on the deduced dynamic graph and the inference data, physical determination judgment data is obtained.

9. The multi-modal spatio-temporal graph-based disaster evolution path inference method according to claim 7, characterized in that, A T-GCN model is called to perform digital simulation prediction on the deduced dynamic graph to obtain digital determination judgment data, including: The T-GCN model is called and tensor data is constructed based on the deduced dynamic graph; The tensor data is input into the T-GCN model, and at least operations including relationship weighting, forward propagation and probability output are performed on the T-GCN model to obtain initial judgment data; The initial judgment data is post-processed to obtain digital determination judgment data.

10. A disaster evolution path deduction system based on multi-modal spatiotemporal graph, characterized in that, including: A static construction module for constructing a deduced static graph based on shopping mall attribute data; A dynamic construction module for initializing the deduced static graph and outputting a deduced dynamic graph through physical simulation; A region classification module for clustering based on the deduced static graph and the deduced dynamic graph to obtain a classified region cluster; A probability prediction module for performing physical prediction and digital prediction based on the classified region cluster and the deduced dynamic graph to obtain spread probability data; An intervention replay module for injecting an intervention instruction into the deduced static graph and performing re-deduction.

Images