Building initial fire spread situation prediction method, system and device based on space-time evolution modeling and medium

Abstract

The application relates to a building initial fire spread situation prediction method, system, equipment and medium based on space-time evolution modeling. The method comprises the following steps: based on the building information model data of a target building, extracting space units and connection relationships and constructing a building topology graph; initializing a graph neural network fire spread prediction model and an evacuation behavior simulation model containing a reinforcement learning agent group based on the topology graph; coupling the two models and conducting adversarial training on the evacuation agent group; obtaining building fire data to input the model for fire state evolution calculation, calling the trained model to simulate personnel evacuation behavior, and outputting a joint prediction situation containing a future space-time fire spread range and a dynamic evacuation path of personnel. The method can overcome the defects of traditional separate simulation methods that ignore the two-way dynamic interaction of fire and personnel behavior, realize coupled space-time evolution prediction of the fire spread and evacuation process, and improve the accuracy and timeliness of building fire emergency situation judgment.

Description

Technical Field

[0001] This invention belongs to the field of fire rescue technology, and in particular relates to a method, system, equipment and medium for predicting the spread of early-stage fires in buildings based on spatiotemporal evolution modeling. Background Technology

[0002] With the continuous development of building safety and emergency management technologies, accurate prediction of building fire dynamics and efficient simulation of personnel evacuation have become key research directions. Currently, there are various fire simulation software based on computational fluid dynamics or empirical models, as well as evacuation simulation tools based on rule-based or social force models, used to assist in building fire protection design and emergency plan development.

[0003] In traditional techniques, fire spread simulation and personnel evacuation simulation are usually treated as two separate processes. First, environmental parameters such as smoke and heat distribution at different times are calculated using a fire model. Then, these parameters are used as fixed background conditions and input into the evacuation model to assess personnel escape routes and time.

[0004] However, the current approach artificially separates the physical processes of fire from human behavioral decision-making, which has significant limitations. In reality, fire spread and evacuation are highly dynamically coupled: the development of the fire changes the evacuation environment in real time, threatening personnel safety; at the same time, the behavior of evacuees (such as opening or closing doors) directly changes ventilation conditions and passageway connectivity, thus affecting the fire's spread path and speed. This two-way dynamic interaction is ignored in traditional separate simulation frameworks, leading to significant deviations between the predicted fire situation and evacuation paths and the actual disaster evolution process. This makes it difficult to provide a reliable, spatiotemporally synchronized joint situation of disaster and evacuation, thereby affecting the accuracy and timeliness of emergency decision-making. Summary of the Invention

[0005] Therefore, it is necessary to provide methods, systems, equipment, and media for predicting the spread of early-stage fires in buildings based on spatiotemporal evolution modeling to address the aforementioned technical issues.

[0006] Firstly, this application provides a method for predicting the initial fire spread pattern in buildings based on spatiotemporal evolution modeling, including: S1. Based on the building information model data of the target building, extract the internal spatial units of the building and their corresponding connectivity relationships to construct a building topology map; S2. Based on the building topology map, initialize the fire spread prediction model and the evacuation behavior simulation model; wherein, the fire spread prediction model is a graph neural network model, used to propagate the fire state on the building topology map; the evacuation behavior simulation model contains an evacuation agent group composed of multiple reinforcement learning agents. S3. Couple the fire spread prediction model and the evacuation behavior simulation model, conduct adversarial training on the evacuation agent group, and after completing the adversarial training, fix the trained agent strategy in the evacuation behavior simulation model to obtain the trained evacuation behavior simulation model. S4. Acquire building fire data and input the building fire data into the fire spread prediction model to calculate the fire state evolution. Based on the evolving fire state, call the trained evacuation behavior simulation model to simulate personnel evacuation behavior and output a joint prediction situation that includes the future spatiotemporal fire spread range and the dynamic evacuation path of personnel.

[0007] In one embodiment, based on the building information model data of the target building, the internal spatial units of the building and their corresponding connectivity relationships are extracted to construct a building topology map, including: S11. Analyze the building information model data, extract the geometric boundaries and attribute information of all enclosed space units, and define each enclosed space unit as a topology graph node. S12. Based on the physical channel information between enclosed space units, establish the adjacency relationship between nodes, and create a topology graph edge for each pair of space unit nodes that are directly connected through doors, windows or ventilation ducts. S13. Assign an initial node feature vector to each node in the topology graph. The initial node feature vector shall include at least the spatial volume, combustible load density, initial temperature and flue gas concentration. S14. Assign an initial edge feature vector to each edge of the topology graph. The initial edge feature vector includes the basic flow coefficient and initial dynamic weight calculated based on the channel physical size. S15. Based on the topology graph nodes, topology graph edges, and initial edge feature vectors, construct the building topology graph.

[0008] In one embodiment, a fire spread prediction model and an evacuation behavior simulation model are coupled, and an adversarial training process is performed on the evacuation agent swarm. After the adversarial training is completed, the trained agent policies in the evacuation behavior simulation model are fixed to obtain a trained evacuation behavior simulation model, including: S21. Based on the building topology map and fire spread prediction model, construct a multi-dimensional state space for each agent in the evacuation agent group, including its own location and local environmental state. S22. Based on the multi-dimensional state space, the current policy network outputs the actions of each agent in the evacuation agent group, and simultaneously samples the perturbation actions of an adversarial fire agent; wherein, the actions of the evacuation agent group include moving and operating the passage door, and the actions of the adversarial fire agent include applying fire intensity perturbation to a specific edge in the building topology graph. S23. Execute the actions of the evacuation agent group and the perturbation actions of the adversarial fire agent in the training environment. Modify the weight of the corresponding edge according to the action of operating the passage door, and enhance the fire propagation coefficient of the corresponding edge according to the action of applying fire intensity perturbation to obtain the updated environmental state. S24. Based on the updated environmental state, calculate the immediate reward obtained by each evacuation agent. The immediate reward is determined by the proximity to the safe exit, the degree of danger of the environment, and the utility of performing the action of operating the passage door. S25. Using the state, action, reward and new state data accumulated in multiple rounds of training, update the policy network parameters of the evacuation agent swarm through the policy gradient algorithm, and update the policy network parameters of the adversarial fire agent by maximizing the negative reward obtained by the evacuation agent swarm. S26. Repeat S22 to S25 until the policy network of the evacuation swarm converges. Fix the policy network parameters of the evacuation swarm at this time to obtain the trained evacuation behavior simulation model.

[0009] In one embodiment, based on a building topology map and a fire spread prediction model, a multi-dimensional state space is constructed for each agent in the evacuation agent swarm, including its own location and local environmental state, comprising: S31. Obtain the fire status features of the node where the agent is located and all its adjacent nodes from the fire spread prediction model, and encode the node where the agent is located to obtain the node location code; wherein, the fire status features include at least temperature and smoke concentration. S32. Aggregate the fire status features of the node where the agent is located and its neighboring nodes to obtain a local environment perception vector; the local environment perception vector is used to characterize the fire threat around the agent. S33. The node position encoding of the agent and the local environment perception vector are concatenated to generate a multi-dimensional state space.

[0010] In one embodiment, actions of evacuation agents and perturbation actions of adversarial fire agents are performed in the training environment. The weights of corresponding edges are modified based on actions of operating passage doors, and the fire propagation coefficients of corresponding edges are enhanced based on actions of applying fire intensity perturbations to obtain updated environmental states, including: S41. Traverse the operation channel door actions executed by the evacuation intelligent agent group. When an intelligent agent performs an open or close operation on the channel door between any two connected nodes, update the weight of the corresponding edge in the building topology graph to the preset open weight or close weight. Among them, the open weight is the normal passage weight, and the close weight is a maximum value used to simulate physical barriers. S42. Apply the fire intensity perturbation applied by the adversarial fire agent, instruct the application of perturbation intensity to the fire propagation coefficient of the selected edge in the building topology graph, and superimpose the perturbation intensity onto the fire propagation coefficient corresponding to the selected edge to obtain the enhanced edge fire propagation coefficient. S43. Based on the building topology map after weight modification and propagation coefficient enhancement, and the new location information of all agents, obtain the updated environment state.

[0011] In one embodiment, building fire data is acquired and input into a fire spread prediction model for fire state evolution calculation. Based on the evolving fire state, a trained evacuation behavior simulation model is invoked to simulate personnel evacuation behavior, outputting a joint prediction of the future spatiotemporal fire spread range and dynamic evacuation paths, including: S51. Obtain building fire data, which includes temperature, smoke concentration, and access control status data from multiple monitoring points; S52. Integrate building fire data and building topology map, update the fire status characteristics of corresponding nodes and the weights of corresponding edges in the map, and generate the initial situation map at the current moment. S53. Starting from the initial situation map, iteratively execute the fire spread prediction model to predict the fire status of each node in the building at multiple consecutive time steps in the future, and generate the spatiotemporal evolution field of fire spread. S54. At each predicted future time step, the fire status of the current step is used as the environmental input. The strategies of all agents in the trained evacuation behavior simulation model are called to deduce the actions of each agent and simulate the dynamic evacuation path of the crowd from the current position. S55. The spatiotemporal evolution field of fire spread is superimposed and simultaneously visualized with dynamic evacuation paths to obtain a joint prediction situation; the joint prediction situation is used to reflect the interaction process between fire and evacuation.

[0012] In one embodiment, starting with an initial situation map, a fire spread prediction model is iteratively executed to predict the fire status of each node within the building at multiple consecutive time steps, generating a spatiotemporal evolution field of fire spread, including: S61. Use the initial situation map as the current input and input it into the fire spread prediction model to obtain the node fire status prediction value for the next time step. S62. Use the predicted node fire status as the new node feature and keep the edge weights of the building topology unchanged to construct a new situation map. S63. Repeat S61 and S62 to predict the node fire status of the preset prediction step size in sequence. S64. Arrange the predicted node fire status values ​​obtained at each time step in chronological order to generate a spatiotemporal evolution field of fire spread. The spatiotemporal evolution field of fire spread includes node index, time step and fire status characteristics.

[0013] Secondly, this application also provides a building initial fire spread prediction system based on spatiotemporal evolution modeling, including: The building topology map construction module is used to extract the internal spatial units and corresponding connectivity relationships of the target building based on the building information model data of the target building, and construct the building topology map. The model initialization module is used to initialize the fire spread prediction model and the evacuation behavior simulation model based on the building topology map. The fire spread prediction model is a graph neural network model used to propagate the fire state on the building topology map. The evacuation behavior simulation model contains a group of evacuation agents composed of multiple reinforcement learning agents. The model strategy adversarial training module is used to couple the fire spread prediction model and the evacuation behavior simulation model, and to conduct adversarial training on the evacuation agent group. After completing the adversarial training, the trained agent strategy in the evacuation behavior simulation model is fixed to obtain the trained evacuation behavior simulation model. The joint situation prediction module is used to acquire building fire data, input the building fire data into the fire spread prediction model to calculate the fire state evolution, and call the trained evacuation behavior simulation model based on the evolving fire state to simulate personnel evacuation behavior, and output a joint situation prediction that includes the future spatiotemporal fire spread range and the dynamic evacuation path of personnel.

[0014] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0015] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0016] The aforementioned method, system, equipment, and medium for predicting the initial fire spread in buildings based on spatiotemporal evolution modeling lay a precise spatial foundation by constructing a building topology map. It couples a graph neural network fire prediction model with a reinforcement learning evacuation simulation model, enabling bidirectional dynamic interaction between the two. Adversarial training enhances the realism of evacuation behavior simulation, real-time fire data fusion ensures the predictions are accurate, and iterative evolution calculations and synchronous visualization generate a spatiotemporally synchronized joint prediction trend. This accurately captures the interaction patterns between fire and evacuation, avoiding the biases of traditional separate simulations. It provides reliable support for building fire protection design, emergency plan development, and fire emergency decision-making, improving the accuracy and timeliness of emergency decisions and facilitating efficient fire prevention and personnel evacuation guidance. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a method for predicting the spread of early-stage building fires based on spatiotemporal evolution modeling in one embodiment. Figure 2 This is a schematic diagram of a building fire spread prediction system based on spatiotemporal evolution modeling in one embodiment. Figure 3 This is a schematic diagram of a computer device in one embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] In one embodiment, reference Figure 1The document presents a flowchart illustrating the method for predicting the initial fire spread in buildings based on spatiotemporal evolution modeling, as provided in this application. This embodiment uses the application of this method to a fire prediction terminal (hereinafter referred to as the terminal) as an example. The fire prediction terminal can be an industrial-grade server, an embedded computing terminal, or a dedicated fire prediction host. It has a built-in processor, memory, data interface, and graphics rendering function. The processor executes the algorithm logic for each step of the method; the memory stores building information model data, model parameters, and training data; the data interface receives monitoring point data and outputs prediction results; and the graphics rendering function visualizes the joint prediction situation. All algorithm logic is implemented using a suitable programming language. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps: S1. Based on the building information model data of the target building, extract the internal spatial units of the building and their corresponding connectivity relationships to construct a building topology map.

[0021] For example, based on the Building Information Modeling (BIM) data of the target building, the internal spatial units and their corresponding connectivity relationships are extracted to construct a building topology map. The fire prediction terminal first obtains the BIM data of the target building through a data interface. This data originates from standardized BIM files generated during the building design phase and contains core information such as all internal spatial structures, component attributes, and connectivity relationships, serving as the foundational data for constructing the building topology map. The fire prediction terminal preprocesses the acquired BIM data, removing redundant non-spatial related data and retaining core information related to fire spread and personnel evacuation, such as enclosed spaces and passageway structures. Then, a dedicated analytical algorithm is used to extract all independent enclosed spatial units within the building, identifying the connectivity structures between these units, such as doors, windows, and ventilation ducts, clarifying the connectivity relationships between the spatial units. Finally, a graph structure is used to map enclosed spatial units as nodes and connectivity structures as edges, constructing a building topology map that accurately reflects the building's spatial layout and connectivity characteristics.

[0022] S2. Based on the building topology map, initialize the fire spread prediction model and the evacuation behavior simulation model.

[0023] The fire spread prediction model is a Graph Neural Network (GNN) model used to propagate the fire state on the building topology map. The evacuation behavior simulation model consists of an evacuation agent swarm composed of multiple reinforcement learning (RL) agents. The fire prediction terminal calls the building topology map stored in memory, reads core parameters such as the number of nodes, the number of edges, and the dimension of node features, and completes the initialization of both models. The fire spread prediction model, built using GNN, captures the feature interactions between nodes and their neighbors through graph convolution operations, achieving accurate propagation and evolution prediction of the fire state on the building topology map. The model inputs are node features and edge features, and the output is the fire state of each node at different times. During the initialization of the evacuation behavior simulation model, the fire prediction terminal determines the size of the evacuation agent swarm based on the building's designed capacity. Each agent corresponds to one person in the building. It is built using a deep reinforcement learning framework, and action decisions are made through a policy network. Initialization uses a random policy, and the action space and state space boundaries of the agents are preset.

[0024] S3. Couple the fire spread prediction model and the evacuation behavior simulation model, and conduct adversarial training on the evacuation agent group. After completing the adversarial training, fix the trained agent strategy in the evacuation behavior simulation model to obtain the trained evacuation behavior simulation model.

[0025] For example, the fire prediction terminal couples the two models through a data interaction interface, constructing a two-way interaction mechanism. The real-time fire state output by the fire spread prediction model serves as the environmental input to the evacuation behavior simulation model. The actions of the agents in the evacuation behavior simulation model are fed back to the fire spread prediction model, influencing the fire spread state. During adversarial training, adversarial fire agents are introduced to form an adversarial relationship with the evacuation agent group. The goal of the evacuation agent group is to learn the optimal evacuation strategy, while the goal of the adversarial fire agents is to apply fire disturbances to increase the difficulty of evacuation. The fire prediction terminal controls the two agent groups to conduct multiple rounds of interactive training, recording the state, actions, and reward data for each round. The gradient descent algorithm is used to update the policy network parameters of both sides until the policy network of the evacuation agent group converges and the policy output stabilizes. At this point, the policy parameters of the evacuation agents are fixed, resulting in a trained model that accurately simulates personnel evacuation behavior.

[0026] S4. Acquire building fire data and input the building fire data into the fire spread prediction model to calculate the fire state evolution. Based on the evolving fire state, call the trained evacuation behavior simulation model to simulate personnel evacuation behavior and output a joint prediction situation that includes the future spatiotemporal fire spread range and the dynamic evacuation path of personnel.

[0027] For example, the fire prediction terminal acquires real-time fire data collected from monitoring points inside the building via a data interface. After denoising and completion preprocessing, the data is input into the fire spread prediction model, replacing the initial fire state parameters and initiating fire state evolution calculations to predict the fire spread over multiple time steps. Simultaneously, at each time step, the fire prediction terminal inputs the current fire state into a trained evacuation behavior simulation model, invokes fixed agent strategies to simulate evacuation actions and paths, and records the positional change trajectory of each agent. Finally, the fire prediction terminal fuses the fire spread evolution data and evacuation path data, overlaying them using visualization technology to generate a joint prediction situation. This clearly reflects the dynamic interaction between the fire and evacuation, providing support for fire emergency decision-making.

[0028] The aforementioned method for predicting the initial fire spread in buildings based on spatiotemporal evolution modeling lays the foundation by constructing a building topology map that accurately represents the spatial characteristics of buildings. It couples a graph neural network fire spread prediction model with a reinforcement learning evacuation behavior simulation model, achieving a two-way dynamic interaction between the physical processes of fire and personnel behavioral decisions. Adversarial training enhances the realism of evacuation behavior simulation, and real-time building fire data fusion ensures the predictions are accurate. Through iterative evolution calculations and synchronous visualization, a spatiotemporally synchronized joint prediction scenario is generated, accurately capturing the interaction patterns between fire and evacuation. This provides a reliable, spatiotemporally synchronized joint scenario of disaster and evacuation, offering dependable support for building fire protection design, emergency plan development, and fire emergency decision-making. It improves the accuracy and timeliness of emergency decisions, facilitating efficient fire prevention and personnel evacuation guidance.

[0029] In an optional embodiment, based on the building information model data of the target building, the internal spatial units of the building and their corresponding connectivity relationships are extracted to construct a building topology map, including the following steps: S11. Analyze the building information model data, extract the geometric boundaries and attribute information of all enclosed space units, and define each enclosed space unit as a topology graph node.

[0030] Optionally, the fire prediction terminal incorporates a dedicated BIM data parsing engine. Employing a layered parsing approach, it separates architectural, structural, and MEP (Mechanical, Electrical, and Plumbing) modules from the BIM data, focusing on extracting data related to enclosed spatial units. Geometric boundary information includes the spatial unit's 3D coordinate range and outline shape, used to define the spatial unit's location and size. Attribute information includes the spatial unit's functional type, combustible material type, and ceiling height, used for subsequent node feature assignment. After parsing, the fire prediction terminal uniquely identifies each enclosed spatial unit, eliminates overlapping and invalid units, maps each independent enclosed spatial unit to a topological graph node, records the correspondence between nodes and spatial units, and establishes a node index table to ensure that each node accurately corresponds to the actual spatial area within the building.

[0031] S12. Based on the physical channel information between enclosed space units, establish the adjacency relationship between nodes, and create a topology graph edge for each pair of space unit nodes that are directly connected through doors, windows or ventilation ducts.

[0032] Optionally, after extracting enclosed space units, the fire prediction terminal further analyzes the passage structure information in the BIM data, filtering out physical structures such as doors, windows, and ventilation ducts that can achieve spatial connectivity, and identifying the two enclosed space units connected by each passage. For doors and windows, only connectivity in the openable state is considered, eliminating fixed components that cannot be opened; for ventilation ducts, ducts that can achieve air circulation and run through two or more space units are selected, ignoring enclosed ducts or those that only connect a single space. Based on the filtered passage information, the fire prediction terminal determines whether any two enclosed space units have a direct connection relationship. If so, a topology graph edge is created between the corresponding two topology graph nodes, the edge identifier is associated with the node identifier, the passage type corresponding to the edge is recorded, and an adjacency table between nodes is established to accurately represent the possible paths for fire spread and personnel evacuation.

[0033] S13. Assign an initial node feature vector to each node in the topology graph.

[0034] The initial node feature vector includes at least the spatial volume, combustible material load density, initial temperature, and smoke concentration. Optionally, the initial node feature vector is the core input parameter of the fire spread prediction model. The fire prediction terminal assigns feature parameters one by one according to the attribute information of the enclosed space unit corresponding to each node. The spatial volume is obtained by analyzing the three-dimensional coordinate range of the enclosed space unit and using geometric calculation methods, representing the size of the space unit; the combustible material load density is preset according to the functional type of the space unit and refers to the building fire protection design code, representing the density of combustible material distribution in the space; the initial temperature is preset to the normal indoor temperature of the building, representing the ambient temperature under fire-free conditions; the initial smoke concentration is preset to the normal air pollutant concentration under fire-free conditions, representing the air quality when there is no fire. All feature parameters are organized in vector form, uniformly normalized, assigned to the corresponding topology graph nodes, and stored in the node feature matrix to ensure the consistency and standardization of the model input data.

[0035] S14. Assign an initial edge feature vector to each edge of the topology graph. The initial edge feature vector includes the basic flow coefficient and initial dynamic weight calculated based on the physical size of the channel.

[0036] Optionally, the initial edge feature vector is used to characterize the hindering or promoting effect of the passage on fire spread and personnel evacuation. The fire prediction terminal calculates and assigns feature parameters based on the passage structure information corresponding to each edge. The basic flow coefficient is calculated based on the passage type and physical dimensions. The basic flow coefficient of doors and windows is positively correlated with the effective cross-sectional area of ​​the passage, while the basic flow coefficient of ventilation ducts is related to the inner diameter and length of the duct, characterizing the flow efficiency of fire smoke and heat within the passage. The initial dynamic weight is preset according to the evacuation priority of the passage. The edge weight corresponding to the evacuation passage is higher than that of ordinary room doors, and ordinary room doors are higher than those of ventilation ducts, characterizing the importance of the passage in personnel evacuation. The larger the weight value, the more significant the impact on evacuation and fire spread. All feature parameters are integrated into the initial edge feature vector, assigned to the corresponding topological graph edges, and stored in the edge feature matrix, which, together with the node feature vectors, supports model calculation. S15. Based on the topology graph nodes, topology graph edges, and initial edge feature vectors, construct the building topology graph.

[0037] Optionally, after completing node definition, adjacency establishment, and assignment of node and edge feature vectors, the fire prediction terminal adopts a graph structure storage format to integrate all relevant data. Specifically, it integrates and associates data such as the topology graph node set, topology graph edge set, node feature matrix, edge feature matrix, and node adjacency table, and organizes them using graph database storage technology to construct a complete building topology graph. This building topology graph fully preserves the target building's spatial structure, connectivity, and core feature parameters related to fire propagation and personnel evacuation, accurately mapping the actual spatial layout and passage characteristics of the building, and is synchronously stored in the fire prediction terminal's memory.

[0038] In an optional embodiment, the fire spread prediction model and the evacuation behavior simulation model are coupled, and the evacuation agent swarm is subjected to adversarial training. After the adversarial training is completed, the trained agent policies in the evacuation behavior simulation model are fixed to obtain the trained evacuation behavior simulation model, including the following steps: S21. Based on the building topology map and fire spread prediction model, construct a multi-dimensional state space for each agent in the evacuation agent swarm, including its own location and local environmental state.

[0039] Optionally, the fire prediction terminal obtains the fire status features of each agent's node and all its neighboring nodes from the fire spread prediction model, covering at least the two core features of temperature and smoke concentration. To avoid the impact of numerical differences in different features on the model training effect, the obtained fire status features are subjected to min-max data normalization to map their numerical range to a reasonable interval. Simultaneously, one-hot encoding is performed on the agent's node to generate a node position code. The dimension of this code is consistent with the total number of nodes in the building topology map. In the encoding vector, only the position corresponding to the agent's node is 1, and the rest are 0, thus achieving accurate positioning of the agent itself. Subsequently, a weighted aggregation algorithm is used to aggregate the fire status features of the agent's node and its neighboring nodes. The aggregation weight is determined based on the distance between the neighboring nodes and the agent's node; the closer the distance, the greater the weight. The weight expression is: In the formula, Representing adjacent nodes Aggregate weights, Representing adjacent nodes The distance to the node where the agent is located can be accurately calculated using the 3D coordinates of the node in the building information model data. After aggregation processing, a local environment perception vector is obtained, which is used to accurately characterize the fire threat situation around the agent. Finally, the node position encoding of the agent is concatenated with the local environment perception vector to generate a multi-dimensional state space, which can comprehensively cover all the information required for the agent's action decision-making.

[0040] S22. Based on the multi-dimensional state space, output the action of the current policy network of each agent in the evacuation agent group, and simultaneously sample the perturbation action of an adversarial fire agent.

[0041] Among them, the actions of the evacuation intelligent agents include moving and manipulating passage doors, while the actions of the fire-resistant intelligent agents include applying fire intensity perturbations to specific edges in the building topology map.

[0042] Optionally, after constructing the multidimensional state space for each agent, the fire prediction terminal controls the evacuation agent group and the adversarial fire agent to execute actions synchronously, achieving interactive adversarial training. For each agent in the evacuation agent group, based on its own constructed multidimensional state space, the execution probability of each possible action is output through the current policy network. Then, a roulette wheel method is used to select the actual action to be executed by each agent according to the action execution probability. This method can ensure the randomness and rationality of action selection and improve the generalization ability of model training.

[0043] Meanwhile, a perturbation action of an adversarial fire agent is selected through random sampling to ensure the diversity of training scenarios and simulate different fire emergencies. The actions of the evacuation agents are strictly divided into two categories: movement and door manipulation. Movement actions involve moving to all neighboring nodes of the agent's current node, with the number of actions matching the number of neighboring nodes. Door manipulation actions include opening and closing doors, allowing for control over the connectivity of passageways. The actions of the adversarial fire agent are strictly limited to applying fire intensity perturbations to specific edges in the building topology graph to simulate sudden escalation during fire spread.

[0044] S23. Execute the actions of the evacuation agent group and the perturbation actions of the adversarial fire agent in the training environment. Modify the weight of the corresponding edge according to the action of operating the passage door, and enhance the fire propagation coefficient of the corresponding edge according to the action of applying fire intensity perturbation to obtain the updated environmental state.

[0045] Optionally, the fire prediction terminal first iterates through all the operation channel door actions performed by the evacuation intelligent agent group. If an agent performs an opening operation on a channel door between any two connected nodes, the weights of the edges corresponding to these two nodes in the building topology diagram are updated to preset opening weights. The opening weights are normal passage weights, which can ensure that the intelligent agents and the fire situation can pass through the channel normally. If an intelligent agent performs a closing operation on a channel door, the weights of the corresponding edges are updated to preset closing weights. The closing weights are set to a maximum value to simulate the physical barrier effect, which can effectively prevent the intelligent agents and the fire situation from passing through the channel, closely matching the actual barrier situation after the door is closed.

[0046] Simultaneously, in response to the fire intensity perturbation action applied by the adversarial fire agent, the fire prediction terminal applies a corresponding perturbation intensity to the fire propagation coefficient of the selected edge in the building topology map according to the perturbation command. This perturbation intensity is then superimposed on the original fire propagation coefficient of that edge to obtain the enhanced edge fire propagation coefficient, expressed as: In the formula This represents the enhanced edge fire propagation coefficient. Represents the original fire propagation coefficient. This represents the intensity of the disturbance. The enhanced fire propagation coefficient will be directly used to adjust the propagation speed of the fire state in this channel. The greater the disturbance intensity, the faster the fire propagation speed, thereby accurately simulating the scenario of escalating fire spread. After the action execution and parameter adjustment are completed, the fire prediction terminal integrates the building topology map after weight modification and propagation coefficient enhancement, the new location information of all agents, the real-time fire status of each node, and other data to finally obtain the updated environmental state. This updated environmental state is stored in the terminal memory in real time to ensure the traceability of the training process.

[0047] S24. Based on the updated environmental state, calculate the immediate reward obtained by each evacuation agent. The immediate reward is determined by the proximity to the safe exit, the degree of danger of the environment, and the utility of performing the action of operating the passage door.

[0048] Optionally, after the updated environmental state is determined, the fire prediction terminal accurately calculates the immediate reward obtained by each evacuation agent after executing the action in this round based on this state. The core of setting the immediate reward is to guide the agent to learn the optimal evacuation strategy. Its calculation strictly combines the core objectives of evacuation behavior and consists of three main parts: the proximity to the safe exit, the degree of danger of the environment, and the effectiveness of executing the action of operating the passage door. It is calculated using a linear weighted summation method, and the expression is: . in the formula Represents an instant reward. , , Represents the weighting coefficient, and strictly satisfies ,in Focusing on evacuation efficiency, Prioritizing personnel safety The focus is on the effectiveness of the action, and the weighting coefficients can be adjusted according to the actual application scenario, but the sum of the three must always be kept at 1. Rewards represent the degree of proximity to the emergency exit. Rewards represent the degree of danger in the environment. This represents the utility reward for performing the operation of the channel door.

[0049] Among them, the key function of the proximity reward to the safe exit is to guide the agent to move towards the safe exit, thereby improving evacuation efficiency. The expression is: In the formula Rewards represent the degree of proximity to the emergency exit. This represents the distance between the agent and the safety exit at the previous moment. This represents the distance between the agent and the safety exit at the current moment. If the distance between the agent's current position and the safety exit is less than that at the previous moment, then... The value is positive, and the greater the decrease in distance, the better. The larger the value, the greater the distance. The value is negative, and its absolute value is proportional to the increase in distance, thus punishing the agent for moving away from the safety exit.

[0050] The core function of the reward based on the degree of danger of the environment is to guide the agent away from dangerous environments and ensure personnel safety. The expression is: In the formula Rewards represent the degree of danger in the environment. Represents the current node temperature. Represents the dangerous temperature threshold. Represents the current flue gas concentration at the node. This represents the threshold for hazardous smoke concentration. If the fire status characteristics of the node where the agent is currently located are within the preset safety threshold range, then... If positive; if it exceeds the safety threshold, then... The value is negative, and the higher the level of danger, The larger the negative value, the more accurately the formula can quantify the impact of environmental hazards on the intelligent agent.

[0051] The core function of the utility reward for performing the action of operating the passage door is to guide the agent to perform effective door operations, thereby achieving fire isolation and optimization of evacuation routes. If the opening operation can provide a better evacuation route for itself or other agents, the utility reward is positive; if the closing operation can prevent the spread of fire and protect the safety of itself or other agents, the utility reward is positive; otherwise, the utility reward is negative. The utility value is quantitatively calculated based on the degree of impact of the action on the evacuation success rate and the spread of fire.

[0052] S25. Using the state, action, reward and new state data accumulated in multiple rounds of training, update the policy network parameters of the evacuation agent swarm through the policy gradient algorithm, and update the policy network parameters of the adversarial fire agent by maximizing the negative reward obtained by the evacuation agent swarm.

[0053] Optionally, after each round of training, the fire prediction terminal stores the state from this round of training, the actions performed by the agents, the immediate rewards obtained by each agent, and the updated environmental state (i.e., the new state) in an experience replay pool. The experience replay pool has a corresponding capacity and updates the data using a first-in, first-out (FIFO) approach, promptly removing outdated data to ensure the timeliness and relevance of the stored data. When a certain amount of multi-round training data accumulates in the experience replay pool, the fire prediction terminal uses this data to synchronously update the policy network parameters of the evacuation agent group and the adversarial fire agent group, achieving adversarial optimization between the two sides.

[0054] Specifically, for the evacuation agent swarm, the fire prediction terminal randomly samples a batch of training data from the experience playback pool, and calculates the target Q-value for each agent using a temporal difference learning algorithm. The expression for the target Q-value is: In the formula Represents the target Q value. Represents an instant reward. This represents a discount factor used to balance the weights of immediate and future rewards. Represents a new state All actions The maximum Q-value is calculated. After obtaining the target Q-value, the mean square error between the Q-value output by the current policy network and the target Q-value is further calculated. This mean square error is used as the loss function, and the weight and bias parameters of the evacuation agent swarm policy network are updated through the backpropagation algorithm. The learning rate is set to a reasonable value to ensure the stability of parameter updates, avoid training oscillations, and gradually optimize the behavior policy of the evacuation agents to improve the rationality of evacuation decisions.

[0055] For adversarial fire agents, the core objective of parameter updates is to achieve adversarial optimization. This is achieved by maximizing the negative reward obtained by the evacuation agent swarm to update its policy network parameters. The loss function of the adversarial fire agent is set as follows: In the formula The loss value represents the damage to the fire-fighting agent. Represents the summation symbol. This represents the sum of immediate rewards for the evacuation agent swarm. The policy network parameters of the adversarial fire agents are updated based on this loss function using the backpropagation algorithm. This allows the adversarial fire agents to gradually learn more targeted perturbation actions, increasing the training difficulty and further optimizing the environmental adaptability and decision-making robustness of the evacuation agent swarm.

[0056] S26. Repeat S22 to S25 until the policy network of the evacuation swarm converges. Fix the policy network parameters of the evacuation swarm at this time to obtain the trained evacuation behavior simulation model.

[0057] Optionally, the fire prediction terminal repeatedly executes the operation process from S22 to S25, namely, action output, environment update, reward calculation, and parameter update, continuously advancing the adversarial training between the evacuation intelligent agent swarm and the adversarial fire intelligent agent. During the training process, the convergence status of the evacuation intelligent agent swarm's policy network is monitored in real time. The convergence criterion is that the agent's evacuation strategy tends to be stable, and after multiple iterations of training, the agent's action decisions no longer show significant optimization, and the cumulative reward tends to be stable. When the policy network convergence is detected, the fire prediction terminal fixes the policy network parameters of the evacuation intelligent agent swarm at this time, stops training, and obtains a trained evacuation behavior simulation model. This model can accurately simulate personnel perception, decision-making, and evacuation behavior based on real-time fire conditions, adapting to the two-way coupling relationship between fire and evacuation.

[0058] In an optional embodiment, based on the building topology map and the fire spread prediction model, a multi-dimensional state space containing the agent's own location and local environmental state is constructed for each agent in the evacuation agent swarm, including the following steps: S31. Obtain the fire status characteristics of the node where the agent is located and all adjacent nodes from the fire spread prediction model, and encode the node where the agent is located to obtain the node location code.

[0059] The fire status characteristics include at least temperature and smoke concentration. The fire prediction terminal first extracts the fire status characteristics of the topological node where each evacuation agent is located and all its adjacent nodes from the initialized fire spread prediction model. These fire status characteristics, including at least temperature and smoke concentration, characterize the fire threat level of the spatial unit corresponding to each node. The initial stage uses the baseline status characteristics from model initialization. Subsequently, a node encoding algorithm is used to uniquely encode the topological node where each evacuation agent is located, generating a node location code. This encoding information accurately characterizes the spatial location of the agent, its associated nodes in the building topology map, and its spatial coordinates, ensuring that the agent can clearly perceive its own spatial location.

[0060] S32. Aggregate the fire status features of the node where the agent is located and its neighboring nodes to obtain a local environment perception vector.

[0061] Optionally, the fire prediction terminal employs a feature aggregation algorithm to aggregate the fire status features (temperature, smoke concentration, etc.) of the node where the agent is located and all its neighboring nodes. During the aggregation process, different weights are assigned to the fire status features of different nodes based on the distance between neighboring nodes and the node where the agent is located, and the weights of their connectivity relationships. The closer the nodes are and the stronger the connectivity, the higher the weight, ensuring that the aggregation results accurately reflect the actual fire threat around the agent. After aggregation, a local environmental perception vector is generated. This vector can centrally represent the distribution of fire status, the degree of fire threat, and the range of dangerous areas within a certain range around the agent, enabling the agent to clearly perceive the safety status of its surrounding environment.

[0062] S33. The node position encoding of the agent and the local environment perception vector are concatenated to generate a multi-dimensional state space.

[0063] Optionally, the fire prediction terminal employs a vector concatenation algorithm to fuse node location codes and local environmental perception vectors, forming a multi-dimensional state space exclusive to each evacuation agent. During the concatenation process, it ensures that the dimensions of the node location codes and local environmental perception vectors match to avoid data redundancy or missing data. The fused multi-dimensional state space simultaneously contains the agent's own location information and fire threat information from the surrounding environment, comprehensively and accurately representing the agent's overall state. This multi-dimensional state space serves as the input to the evacuation agent's strategy network, directly influencing the agent's action output and decision-making logic, ensuring that the agent can make reasonable evacuation decisions based on its own location and the surrounding fire situation.

[0064] In an optional embodiment, the actions of the evacuation agent swarm and the perturbation actions of the adversarial fire agent are executed in the training environment. The weights of the corresponding edges are modified according to the actions of operating the passage doors, and the fire propagation coefficients of the corresponding edges are enhanced according to the actions of applying fire intensity perturbations to obtain the updated environmental state, including the following steps: S41. Traverse the operation channel door actions executed by the evacuation intelligent agent group. When an agent performs an open or close operation on the channel door between any two connected nodes, update the weight of the corresponding edge in the building topology graph to the preset open or close weight.

[0065] Optionally, the fire prediction terminal iterates through all the operation door actions performed by the evacuation intelligent agent group, matching each action with the corresponding building topology edge (i.e., the connected edge corresponding to the door). When an agent performs the action of opening a door, the weight of the corresponding edge is updated to a preset opening weight, which is a normal passage weight used to characterize the connectivity of personnel and smoke / heat passing through normally when the door is open. When an agent performs the action of closing a door, the weight of the corresponding edge is updated to a preset closing weight, which is set to a maximum value to simulate the physical barrier effect after the door is closed, i.e., neither fire nor personnel can pass through the passage, thus realizing the dynamic update of the passage connectivity status.

[0066] S42. Apply the fire intensity perturbation applied by the adversarial fire agent to indicate the perturbation intensity applied to the fire propagation coefficient of the selected edge in the building topology graph, and superimpose the perturbation intensity onto the fire propagation coefficient corresponding to the selected edge to obtain the enhanced edge fire propagation coefficient.

[0067] Optionally, after updating the edge weights corresponding to the passageway door operation, the fire prediction terminal simultaneously applies a fire intensity perturbation action sampled by the adversarial fire agent. This perturbation action explicitly indicates the selected edge in the building topology map that needs to be perturbed, as well as the corresponding perturbation intensity. Based on this indication, the fire prediction terminal accurately locates the selected edge in the topology map and extracts the current fire propagation coefficient corresponding to that edge. Subsequently, the perturbation intensity is superimposed on the original fire propagation coefficient of the selected edge, and the enhanced edge fire propagation coefficient is obtained through superposition calculation. The greater the perturbation intensity, the faster the fire spreads in the passageway, thereby accurately simulating the sudden escalation scenario during the fire spread process.

[0068] S43. Based on the building topology map after weight modification and propagation coefficient enhancement, and the new location information of all agents, obtain the updated environment state.

[0069] Optionally, after modifying the weights of corresponding edges in the building topology map and enhancing the fire propagation coefficient of selected edges, the fire prediction terminal comprehensively integrates the relevant data. On the one hand, it retains the complete building topology map after weight modification and propagation coefficient enhancement, which accurately reflects the latest changes in channel connectivity and fire propagation efficiency. On the other hand, it synchronously collects the new location information of all evacuation agents after performing actions, including the updated node positions after performing movement actions and the current node positions maintained after performing actions to operate channel doors, ensuring the accuracy and real-time nature of agent position information. Finally, it deeply integrates the modified building topology map with the new location information of all agents to form the updated environment state after the execution of this round of actions, and stores this updated environment state in the terminal memory in real time, ensuring the traceability of the training process.

[0070] In an optional embodiment, building fire data is acquired and input into a fire spread prediction model for fire state evolution calculation. Based on the evolving fire state, a trained evacuation behavior simulation model is invoked to simulate personnel evacuation behavior, outputting a joint prediction of the future spatiotemporal fire spread range and dynamic evacuation paths, including: S51. Obtain building fire data.

[0071] The building fire data includes temperature, smoke concentration, and access control status data from multiple monitoring points.

[0072] Optionally, the fire prediction terminal acquires real-time building fire data through its own data interface. This fire data originates from a dedicated fire monitoring system deployed within the building, ensuring the data's real-time accuracy, authenticity, and completeness. The deployment locations of monitoring points precisely correspond to nodes in the building's topology map. Each monitoring point is equipped with high-precision sensing devices: a temperature sensor for real-time acquisition of the ambient temperature in the corresponding node area, a smoke sensor for real-time acquisition of the smoke concentration in the corresponding node area, and an access control sensor for real-time acquisition of the opening and closing status of the corresponding passageway doors. These three sensors together constitute complete building fire data. The monitoring data is uploaded to the fire prediction terminal in real-time via wireless transmission at an appropriate frequency to ensure timely capture of dynamic changes in the fire scene.

[0073] S52. Integrate building fire data and building topology map, update the fire status characteristics of corresponding nodes and the weights of corresponding edges in the map, and generate the initial situation map at the current moment.

[0074] Optionally, to ensure the compatibility of fire scene data with building topology maps, the fire prediction terminal first preprocesses the acquired building fire scene data to eliminate data noise and anomalies, ensuring data accuracy: data denoising uses a moving average filtering algorithm, sets a corresponding window size, and replaces the current time data with the average of monitoring data from multiple adjacent time points, effectively eliminating random noise; outlier removal uses reasonable criteria to remove data that exceeds the sensor measurement range or deviates from the normal fluctuation range; for missing data, a linear interpolation algorithm is used to complete it, accurately calculating the missing value based on the valid data of adjacent time points before and after the missing data, ensuring data continuity.

[0075] After preprocessing, the building fire data and the building topology map are deeply fused. Based on the one-to-one correspondence between monitoring point identifiers and topology map node identifiers, the real-time temperature and smoke concentration data of each monitoring point are updated to the fire status characteristics of the corresponding node, replacing the default values ​​during initialization to ensure that the node fire status is consistent with the actual fire scene. The access control status data of each monitoring point is updated to the weights of the corresponding edges. If the access control status is closed, the weight of the corresponding edge is updated to the closed weight; if it is open, it is updated to the open weight, ensuring that the edge weights are consistent with the actual channel connectivity status. Finally, the fusion generates the initial situation map for the current moment, which serves as the starting point for subsequent fire status evolution calculations and is synchronously stored in the terminal memory.

[0076] S53. Starting from the initial situation map, iteratively execute the fire spread prediction model to predict the fire status of each node in the building at multiple consecutive time steps in the future, and generate the spatiotemporal evolution field of fire spread.

[0077] Optionally, the fire prediction terminal uses the generated initial situation map as the first input, iteratively executes the fire spread prediction model initialized by S2, and gradually predicts the fire status of each node in the building at multiple consecutive time steps in the future, ultimately forming a spatiotemporal evolution field of fire spread, accurately reflecting the dynamic trend of future fire spread. The specific iteration process is as follows: First, the initial situation map is input into the fire spread prediction model. The model uses its internal graph convolutional layer to accurately extract and propagate node and edge features. Combining the propagation coefficient set by S2 and the updated edge fire propagation coefficient, the predicted fire status value of each node in the next time step is calculated. The time step is set to a reasonable value, balancing prediction accuracy and computational efficiency. The predicted fire status value strictly includes four core parameters: temperature, smoke concentration, fire intensity, and combustion rate. Each parameter prediction uses a targeted correction model to ensure prediction accuracy: temperature prediction uses a correction model based on the heat conduction equation, the expression of which is: In the formula, This represents the predicted node temperature for the next time step. Represents the current node temperature. The fire propagation coefficient representing the edge, Represents the average temperature of adjacent nodes. Represents the rate of heat generation from the combustion of combustibles. The time step is represented by the value; the flue gas concentration prediction uses a model based on the flue gas diffusion equation, expressed as: In the formula, This represents the predicted flue gas concentration for the next time step. This represents the current flue gas concentration at the node. Represents the flue gas diffusion coefficient. This represents the average flue gas concentration at adjacent nodes. This represents the smoke generation rate. Subsequently, the predicted node fire state values ​​are used as new node features, and the node feature vectors are updated synchronously while keeping the edge weights of the building topology unchanged (edge ​​weights are only updated when the agent performs an operation on a door or the access control status changes during fire evolution). A new situation map is then constructed as the input for the next time step. This prediction and situation map construction process is repeated to predict the node fire states for the preset prediction steps in sequence. Finally, the predicted node fire states for each time step are arranged chronologically to generate a spatiotemporal evolution field of fire spread. This spatiotemporal evolution field is stored in a three-dimensional array with dimensions of [prediction step length × total number of nodes × number of fire state features], clearly containing the three core pieces of information: node index, time step, and fire state features.

[0078] S54. At each predicted future time step, the fire status of the current step is used as the environmental input. The strategies of all agents in the trained evacuation behavior simulation model are invoked to deduce the actions of each agent and simulate the dynamic evacuation path of the crowd starting from the current position.

[0079] Optionally, at each predicted future time step, the fire prediction terminal uses the fire state corresponding to that time step (i.e., the fire state characteristics of the nodes in the current step situation diagram) as environmental input, calls the fixed strategies of all agents in the trained evacuation behavior simulation model, deduce the optimal action of each agent, and then accurately simulates the dynamic evacuation path of the crowd from the current position, so as to realize the synchronous simulation of fire spread and personnel evacuation.

[0080] Specifically, for the situation map corresponding to each time step, the fire prediction terminal first extracts the fire state features of each node and synchronously inputs them into each evacuation agent to ensure that the agent can perceive the dynamic changes of the current fire environment in real time. Each agent, based on its current position and local environmental state, outputs the execution probability of each action through a trained policy network and selects the action with the highest probability to ensure optimal action decision-making. If an agent performs a movement action, its position is synchronously updated to the corresponding adjacent node, and the fire prediction terminal records the position change information of each agent in real time. If an agent performs an operation action on the passage door, the weight of the corresponding edge in the building topology map is synchronously updated, and the action is fed back to the fire spread prediction model in real time for fire state prediction in subsequent time steps, further strengthening the dynamic coupling of the two models.

[0081] The fire prediction terminal connects the location information of each agent at each time step in chronological order to form the individual dynamic evacuation path of each agent. The paths of all agents together constitute the evacuation path distribution of the population, taking into account both the overall trend and individual differences.

[0082] S55. The spatiotemporal evolution field of fire spread is superimposed and simultaneously visualized with dynamic evacuation paths to obtain a joint prediction situation.

[0083] Optionally, after generating the spatiotemporal evolution field of fire spread and simulating the dynamic evacuation paths of personnel, the fire prediction terminal overlays and integrates the two and performs synchronous visualization processing to obtain a joint prediction situation, realizing an integrated presentation of the interaction between fire and evacuation. The specific visualization process is as follows: using the terminal's built-in graphics rendering module, the building topology map is used as the underlying background, and differentiated color gradients are used to mark the fire spread range at different time steps. The darker the color, the more intense the fire, clearly presenting the spatiotemporal evolution of the fire. At the same time, lines of different colors are used to mark the dynamic evacuation paths of each agent. The thickness of the lines corresponds to the number of agents on the path. The thicker the line, the more people are evacuated on the path, intuitively reflecting the overall trend of personnel evacuation and individual path differences.

[0084] This overlay visualization method allows the joint forecast situation to clearly and intuitively reflect the dynamic interaction between fire spread and personnel evacuation, clarifying the impact of fire on evacuation at different time points and the indirect effect of evacuation behavior on fire spread. The joint forecast situation has multiple output methods: it can be displayed in real time on the fire prediction terminal for on-site personnel to view; it can also be output to the fire emergency command system via a data interface, providing accurate and intuitive support for firefighters to make emergency decisions, effectively improving the accuracy and timeliness of emergency rescue.

[0085] In an optional embodiment, starting from an initial situation map, a fire spread prediction model is iteratively executed to predict the fire status of each node within the building at multiple consecutive time steps, generating a spatiotemporal evolution field of fire spread, including: S61. The initial situation map is used as the current input and input into the fire spread prediction model to obtain the node fire status prediction value for the next time step.

[0086] Optionally, the fire prediction terminal uses the generated initial situation map (containing updated node features and edge weights) as the current input data and inputs it into the trained fire spread prediction model (graph neural network model). The model uses a node feature propagation algorithm to propagate the current fire state features of each node along the edges of the topology graph. At the same time, it uses a feature aggregation algorithm to aggregate the fire state features of adjacent nodes. Combined with the fire propagation coefficient of the edges, it calculates the predicted fire state of all topology nodes in the building at the next time step. This predicted value includes key parameters such as temperature and smoke concentration of each node, which are used to characterize the fire state of each spatial unit at the next moment.

[0087] S62. Use the predicted node fire status as the new node feature and keep the edge weights of the building topology unchanged to construct a new situation map.

[0088] Optionally, the fire prediction terminal updates the predicted fire status of a node for the next time step with the new features of the corresponding node in the building topology graph, replacing the fire status features of that node in the previous time step, thus achieving dynamic updating of node features. Simultaneously, considering that the state of passageway doors does not change frequently in a short period, the edge weights of the building topology graph remain unchanged to ensure the stability of passageway connectivity. Subsequently, the updated node features and the unchanged edge weights are integrated to construct a new situation map, which corresponds to the building fire status and spatial connectivity status for the next time step, providing input data for the next round of iterative prediction.

[0089] S63. Repeat S61 and S62 to predict the node fire status of the preset prediction step size in sequence.

[0090] Optionally, the fire prediction terminal repeatedly executes the S61 and S62 operation procedures. That is, using the new situation map constructed in the previous time step as input, the fire spread prediction model calculates the predicted node fire status value for the next time step, then updates node features, constructs a new situation map, and iterates sequentially. During the iteration process, it strictly adheres to the preset number of prediction steps, which is set according to actual fire emergency needs to ensure coverage of the critical development stages in the early stages of a fire. Each iteration achieves fire status prediction for one time step, gradually obtaining the fire status of all topological nodes within the building at each time step within the preset prediction period, fully capturing the dynamic evolution of the future fire.

[0091] S64. Arrange the predicted node fire status values ​​obtained at each time step in chronological order to generate a spatiotemporal evolution field of fire spread. The spatiotemporal evolution field of fire spread includes node index, time step and fire status characteristics.

[0092] Optionally, the fire prediction terminal collects the predicted fire status values ​​of all time steps obtained from iterative prediction, organizes and arranges them in chronological order, and constructs a spatiotemporal evolution field of fire spread. This spatiotemporal evolution field contains three core elements: node index, time step, and fire status characteristics. The spatiotemporal evolution field can completely and clearly characterize the spread path, spread speed, and fire status change patterns of the fire within the building over a future period, achieving a spatiotemporal synchronous presentation of the fire spread situation. The node index corresponds to each node in the building topology map and is used to identify the spatial unit corresponding to the node. The time step is used to identify the time node corresponding to each fire status. The fire status characteristics include parameters such as temperature and smoke concentration for each node at the corresponding time step.

[0093] The aforementioned method for predicting the initial fire spread in buildings based on spatiotemporal evolution modeling lays the foundation by constructing a building topology map that accurately represents the spatial characteristics of buildings. It couples a graph neural network fire spread prediction model with a reinforcement learning evacuation behavior simulation model, achieving a two-way dynamic interaction between the physical processes of fire and personnel behavioral decisions. Adversarial training enhances the realism of evacuation behavior simulation, and real-time fire scene data fusion ensures the prediction closely matches the actual scenario. Through iterative evolution calculations and synchronous visualization, a spatiotemporally synchronized joint prediction trend is generated, avoiding the biases of traditional separate simulations. This method accurately captures the interaction patterns between fire and evacuation, providing reliable support for building fire protection design, emergency plan formulation, and fire emergency decision-making. It improves the accuracy and timeliness of emergency decisions, assists in the efficient implementation of fire prevention and personnel evacuation guidance, and ensures building safety and the safety of human lives.

[0094] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0095] Based on the same inventive concept, this application also provides a system for predicting the spread of an initial building fire based on spatiotemporal evolution modeling, used to implement the aforementioned method for predicting the spread of an initial building fire based on spatiotemporal evolution modeling. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the system for predicting the spread of an initial building fire based on spatiotemporal evolution modeling provided below can be found in the limitations of the method for predicting the spread of an initial building fire based on spatiotemporal evolution modeling described above, and will not be repeated here.

[0096] In one exemplary embodiment, such as Figure 2 As shown, it presents a structural schematic diagram of a building initial fire spread prediction system 10 based on spatiotemporal evolution modeling, including: Building topology construction module 11 is used to extract the internal spatial units and corresponding connectivity relationships of the building based on the building information model data of the target building, and construct the building topology map. The model initialization module 12 is used to initialize the fire spread prediction model and the evacuation behavior simulation model based on the building topology map. The fire spread prediction model is a graph neural network model used to propagate the fire state on the building topology map. The evacuation behavior simulation model contains a group of evacuation agents composed of multiple reinforcement learning agents. The model strategy adversarial training module 13 is used to couple the fire spread prediction model and the evacuation behavior simulation model, and to conduct adversarial training on the evacuation intelligent agent group. After completing the adversarial training, the trained intelligent agent strategy in the evacuation behavior simulation model is fixed to obtain the trained evacuation behavior simulation model. The joint situation prediction module 14 is used to acquire building fire data, input the building fire data into the fire spread prediction model to calculate the fire state evolution, and call the trained evacuation behavior simulation model based on the evolving fire state to simulate personnel evacuation behavior, and output a joint prediction situation including the future spatiotemporal fire spread range and the dynamic evacuation path of personnel.

[0097] Furthermore, the building topology diagram construction module 11 is also used for: S11. Analyze the building information model data, extract the geometric boundaries and attribute information of all enclosed space units, and define each enclosed space unit as a topology graph node. S12. Based on the physical channel information between enclosed space units, establish the adjacency relationship between nodes, and create a topology graph edge for each pair of space unit nodes that are directly connected through doors, windows or ventilation ducts. S13. Assign an initial node feature vector to each node in the topology graph. The initial node feature vector shall include at least the spatial volume, combustible load density, initial temperature and flue gas concentration. S14. Assign an initial edge feature vector to each edge of the topology graph. The initial edge feature vector includes the basic flow coefficient and initial dynamic weight calculated based on the channel physical size. S15. Based on the topology graph nodes, topology graph edges, and initial edge feature vectors, construct the building topology graph.

[0098] Furthermore, the model policy adversarial training module 13 is also used for: S21. Based on the building topology map and fire spread prediction model, construct a multi-dimensional state space for each agent in the evacuation agent group, including its own location and local environmental state. S22. Based on the multi-dimensional state space, the current policy network outputs the actions of each agent in the evacuation agent group, and simultaneously samples the perturbation actions of an adversarial fire agent; wherein, the actions of the evacuation agent group include moving and operating the passage door, and the actions of the adversarial fire agent include applying fire intensity perturbation to a specific edge in the building topology graph. S23. Execute the actions of the evacuation agent group and the perturbation actions of the adversarial fire agent in the training environment. Modify the weight of the corresponding edge according to the action of operating the passage door, and enhance the fire propagation coefficient of the corresponding edge according to the action of applying fire intensity perturbation to obtain the updated environmental state. S24. Based on the updated environmental state, calculate the immediate reward obtained by each evacuation agent. The immediate reward is determined by the proximity to the safe exit, the degree of danger of the environment, and the utility of performing the action of operating the passage door. S25. Using the state, action, reward and new state data accumulated in multiple rounds of training, update the policy network parameters of the evacuation agent swarm through the policy gradient algorithm, and update the policy network parameters of the adversarial fire agent by maximizing the negative reward obtained by the evacuation agent swarm. S26. Repeat S22 to S25 until the policy network of the evacuation swarm converges. Fix the policy network parameters of the evacuation swarm at this time to obtain the trained evacuation behavior simulation model.

[0099] Furthermore, the model policy adversarial training module 13 is also used for: S31. Obtain the fire status features of the node where the agent is located and all its adjacent nodes from the fire spread prediction model, and encode the node where the agent is located to obtain the node location code; wherein, the fire status features include at least temperature and smoke concentration. S32. Aggregate the fire status features of the node where the agent is located and its neighboring nodes to obtain a local environment perception vector; the local environment perception vector is used to characterize the fire threat around the agent. S33. The node position encoding of the agent and the local environment perception vector are concatenated to generate a multi-dimensional state space.

[0100] Furthermore, the model policy adversarial training module 13 is also used for: S41. Traverse the operation channel door actions executed by the evacuation intelligent agent group. When an intelligent agent performs an open or close operation on the channel door between any two connected nodes, update the weight of the corresponding edge in the building topology graph to the preset open weight or close weight. Among them, the open weight is the normal passage weight, and the close weight is a maximum value used to simulate physical barriers. S42. Apply the fire intensity perturbation applied by the adversarial fire agent, instruct the application of perturbation intensity to the fire propagation coefficient of the selected edge in the building topology graph, and superimpose the perturbation intensity onto the fire propagation coefficient corresponding to the selected edge to obtain the enhanced edge fire propagation coefficient. S43. Based on the building topology map after weight modification and propagation coefficient enhancement, and the new location information of all agents, obtain the updated environment state.

[0101] Furthermore, the joint situation prediction module 14 is also used for: S51. Obtain building fire data, which includes temperature, smoke concentration, and access control status data from multiple monitoring points; S52. Integrate building fire data and building topology map, update the fire status characteristics of corresponding nodes and the weights of corresponding edges in the map, and generate the initial situation map at the current moment. S53. Starting from the initial situation map, iteratively execute the fire spread prediction model to predict the fire status of each node in the building at multiple consecutive time steps in the future, and generate the spatiotemporal evolution field of fire spread. S54. At each predicted future time step, the fire status of the current step is used as the environmental input. The strategies of all agents in the trained evacuation behavior simulation model are called to deduce the actions of each agent and simulate the dynamic evacuation path of the crowd from the current position. S55. The spatiotemporal evolution field of fire spread is superimposed and simultaneously visualized with dynamic evacuation paths to obtain a joint prediction situation; the joint prediction situation is used to reflect the interaction process between fire and evacuation.

[0102] Furthermore, the joint situation prediction module 14 is also used for: S61. Use the initial situation map as the current input and input it into the fire spread prediction model to obtain the node fire status prediction value for the next time step. S62. Use the predicted node fire status as the new node feature and keep the edge weights of the building topology unchanged to construct a new situation map. S63. Repeat S61 and S62 to predict the node fire status of the preset prediction step size in sequence. S64. Arrange the predicted node fire status values ​​obtained at each time step in chronological order to generate a spatiotemporal evolution field of fire spread. The spatiotemporal evolution field of fire spread includes node index, time step and fire status characteristics.

[0103] In one embodiment, such as Figure 3 A computer device 300 is provided, comprising: At least one processor 301, and at least one memory 302 communicatively connected to said processor 301; said memory stores application code executable by said processor, said application code being executed by said processor to enable said processor to perform the steps of the building initial fire spread trend prediction method based on spatiotemporal evolution modeling as described above; The computer device may also include: sensor 303; The processor 301, memory 302, and sensor 303 can be connected via bus 304 or other means. The figure shows an example of connection via bus 304. Figure 3 The character is represented by a single thick line, but this does not mean that there is only one bus or a type of bus.

[0104] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0105] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0106] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for predicting the initial fire spread in buildings based on spatiotemporal evolution modeling, characterized in that, The method includes: S1. Based on the building information model data of the target building, extract the internal spatial units of the building and their corresponding connectivity relationships to construct a building topology map; S2. Based on the building topology map, initialize the fire spread prediction model and the evacuation behavior simulation model; wherein, the fire spread prediction model is a graph neural network model, used to propagate the fire state on the building topology map; the evacuation behavior simulation model includes an evacuation agent group composed of multiple reinforcement learning agents; S3. Couple the fire spread prediction model and the evacuation behavior simulation model, perform adversarial training on the evacuation agent group, and after completing the adversarial training, fix the trained agent strategy in the evacuation behavior simulation model to obtain the trained evacuation behavior simulation model. S4. Acquire building fire data and input the building fire data into the fire spread prediction model to calculate the fire state evolution. Based on the evolving fire state, call the trained evacuation behavior simulation model to simulate personnel evacuation behavior and output a joint prediction situation including the future spatiotemporal fire spread range and the dynamic evacuation path of personnel.

2. The method according to claim 1, characterized in that, The building information model data based on the target building is used to extract the internal spatial units of the building and their corresponding connectivity relationships to construct a building topology map, including: S11. Analyze the building information model data, extract the geometric boundaries and attribute information of all enclosed space units, and define each enclosed space unit as a topology graph node; S12. Based on the physical channel information between the enclosed space units, establish the adjacency relationship between nodes, and create a topology graph edge for each pair of space unit nodes that are directly connected through doors, windows or ventilation ducts. S13. Assign an initial node feature vector to each node of the topology graph. The initial node feature vector includes at least spatial volume, combustible load density, initial temperature and flue gas concentration. S14. Assign an initial edge feature vector to each edge of the topology graph. The initial edge feature vector includes a basic flow coefficient and an initial dynamic weight calculated based on the channel physical size. S15. Based on the topology graph nodes, the topology graph edges, and the initial edge feature vectors, the building topology graph is constructed.

3. The method according to claim 2, characterized in that, The process of coupling the fire spread prediction model and the evacuation behavior simulation model, performing adversarial training on the evacuation agent swarm, and fixing the trained agent policies in the evacuation behavior simulation model after completing the adversarial training to obtain the trained evacuation behavior simulation model includes: S21. Based on the building topology map and the fire spread prediction model, construct a multi-dimensional state space for each agent in the evacuation agent group, including its own position and local environmental state. S22. Based on the multi-dimensional state space, the actions of each agent in the evacuation agent group are output through the current policy network, and the perturbation actions of an adversarial fire agent are sampled at the same time; wherein, the actions of the evacuation agent group include moving and operating the passage door, and the actions of the adversarial fire agent include applying fire intensity perturbation to a specific edge in the building topology graph. S23. Execute the actions of the evacuation intelligent agent group and the disturbance actions of the adversarial fire intelligent agent in the training environment, modify the weight of the corresponding side according to the action of operating the passage door, and enhance the fire propagation coefficient of the corresponding side according to the action of applying fire intensity disturbance to obtain the updated environmental state. S24. Based on the updated environmental state, calculate the immediate reward obtained by each evacuation agent. The immediate reward is determined by the proximity to the safe exit, the degree of danger of the environment, and the utility of performing the operation of the access door. S25. Using the state, action, reward and new state data accumulated in multiple rounds of training, update the policy network parameters of the evacuation agent group through the policy gradient algorithm, and update the policy network parameters of the adversarial fire agent by maximizing the negative reward obtained by the evacuation agent group. S26. Repeat S22 to S25 until the policy network of the evacuation intelligent agent group converges. Fix the policy network parameters of the evacuation intelligent agent group at this time to obtain the trained evacuation behavior simulation model.

4. The method according to claim 3, characterized in that, Based on the building topology map and the fire spread prediction model, a multi-dimensional state space is constructed for each agent in the evacuation agent swarm, including its own location and local environmental state, including: S31. Obtain the fire status features of the node where the agent is located and all adjacent nodes from the fire spread prediction model, and encode the node where the agent is located to obtain the node location code; wherein, the fire status features include at least temperature and smoke concentration. S32. Aggregate the fire status features of the node where the agent is located and its adjacent nodes to obtain a local environment perception vector; the local environment perception vector is used to characterize the fire threat around the agent. S33. The node position encoding of the agent and the local environment perception vector are concatenated to generate the multidimensional state space.

5. The method according to claim 3, characterized in that, The process of executing the actions of the evacuation agent swarm and the perturbation actions of the adversarial fire agent in the training environment, modifying the weights of corresponding edges based on the action of operating the passage door, and enhancing the fire propagation coefficient of corresponding edges based on the action of applying fire intensity perturbation, to obtain an updated environmental state, includes: S41. Iterate through the operation channel door actions performed by the evacuation intelligent agent group. When the intelligent agent performs an open or close operation on the channel door of any two connected nodes, update the weight of the corresponding edge in the building topology graph to a preset open weight or close weight. The open weight is a normal passage weight, and the close weight is a maximum value used to simulate physical obstruction. S42. Apply the fire intensity perturbation applied by the adversarial fire agent to instruct the application of perturbation intensity to the fire propagation coefficient of the selected edge in the building topology graph, and superimpose the perturbation intensity onto the fire propagation coefficient corresponding to the selected edge to obtain the enhanced edge fire propagation coefficient. S43. Based on the building topology map after weight modification and propagation coefficient enhancement, and the new position information of all agents, the updated environment state is obtained.

6. The method according to claim 1, characterized in that, The process involves acquiring building fire data, inputting this data into the fire spread prediction model for fire state evolution calculation, and then using the trained evacuation behavior simulation model to simulate personnel evacuation behavior based on the evolving fire state. The result is a joint prediction of the future spatiotemporal fire spread range and dynamic evacuation paths, including: S51. Obtain building fire data, which includes temperature, smoke concentration and access control status data from multiple monitoring points; S52. The building fire data and the building topology map are fused together, the fire status characteristics of the corresponding nodes and the weights of the corresponding edges in the map are updated, and the initial situation map at the current moment is generated. S53. Starting from the initial situation map, iteratively execute the fire spread prediction model to predict the fire status of each node in the building at multiple consecutive time steps in the future, and generate the spatiotemporal evolution field of fire spread. S54. At each predicted future time step, the fire status of the current step is used as the environmental input. The strategies of all agents in the trained evacuation behavior simulation model are called to deduce the actions of each agent and simulate the dynamic evacuation path of the crowd from the current position. S55. The spatiotemporal evolution field of the fire spread is superimposed and synchronously visualized with the dynamic evacuation path to obtain the joint prediction situation; the joint prediction situation is used to reflect the interaction process between the fire situation and evacuation.

7. The method according to claim 6, characterized in that, Starting from the initial situation map, the fire spread prediction model is iteratively executed to predict the fire status of each node within the building at multiple consecutive time steps, generating a spatiotemporal evolution field of fire spread, including: S61. The initial situation map is used as the current input and input into the fire spread prediction model to obtain the node fire status prediction value for the next time step. S62. Use the predicted node fire status as the new node feature, and keep the edge weights of the building topology unchanged to construct a new situation map. S63. Repeat S61 and S62 to predict the node fire status of the preset prediction step size in sequence. S64. Arrange the predicted node fire status values ​​obtained at each time step in chronological order to generate the spatiotemporal evolution field of the fire spread. The spatiotemporal evolution field of the fire spread includes node index, time step, and fire status characteristics.

8. A building fire spread prediction system based on spatiotemporal evolution modeling, characterized in that, The system includes: The building topology map construction module is used to extract the internal spatial units and corresponding connectivity relationships of the target building based on the building information model data of the target building, and construct the building topology map. The model initialization module is used to initialize a fire spread prediction model and an evacuation behavior simulation model based on the building topology map; wherein, the fire spread prediction model is a graph neural network model used to propagate the fire state on the building topology map; the evacuation behavior simulation model includes an evacuation agent group composed of multiple reinforcement learning agents. The model strategy adversarial training module is used to couple the fire spread prediction model and the evacuation behavior simulation model, and to perform adversarial training on the evacuation agent group. After completing the adversarial training, the trained agent strategy in the evacuation behavior simulation model is fixed to obtain the trained evacuation behavior simulation model. The joint situation prediction module is used to acquire building fire data, input the building fire data into the fire spread prediction model to calculate the fire state evolution, and call the trained evacuation behavior simulation model based on the evolving fire state to simulate personnel evacuation behavior, and output a joint situation prediction that includes the future spatiotemporal fire spread range and the dynamic evacuation path of personnel.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images