Fire-fighting emergency evacuation guiding system driven by reinforcement learning

Abstract

The invention discloses a fire-fighting emergency evacuation guiding system driven by reinforcement learning, relates to the technical field of emergency evacuation guiding, and provides the following scheme that the fire-fighting emergency evacuation guiding system comprises the steps of gridding a monitoring area and calculating a comprehensive danger value of each grid by collecting temperature, smoke concentration, personnel density and a standard path in a building in real time; after the high-risk grids are removed, the remaining grid risk values are input into a reinforcement learning decision model to generate candidate paths; a safety score is obtained by calculating the matching degree of a path and a standard path, and the score is dynamically adjusted in combination with the danger value of each grid and the personnel density; meanwhile, obtaining an efficiency score based on a path length ratio and performing corresponding optimization; and finally, determining an optimal evacuation path from the candidate paths by adopting a multi-objective optimization algorithm, and driving an intelligent indicating device to guide evacuation. According to the invention, the limitation of a fixed evacuation scheme in a dynamic fire scene is solved, intelligent emergency evacuation command is realized, and the safety and efficiency of the evacuation process are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of emergency evacuation guidance, more particularly, the present application relates to a reinforcement learning driven fire emergency evacuation guidance system. BACKGROUND

[0002] With the acceleration of urbanization and the popularity of high-rise buildings, building fire safety protection problems are increasingly prominent. In emergency situations such as fire, how to provide rapid and safe evacuation guidance for people in the building has become an important issue to protect life safety. At present, various buildings are generally equipped with emergency evacuation indication systems, which play an important role in fire emergency evacuation.

[0003] The existing emergency evacuation guidance system mainly adopts a static guidance scheme based on preset rules. The method is to install fixed evacuation indication signs and emergency lighting equipment in the building, and set fixed evacuation directions according to the evacuation plan in the building design stage. The advantage of this prior art is that the system structure is simple, the cost is low, and it can provide basic evacuation guidance under normal circumstances.

[0004] However, the guidance path of the existing static guidance system is a fixed scheme preset in advance, which often cannot be dynamically adjusted according to the development situation of the fire. When the preset path is blocked by flames and smoke, the system cannot automatically re-plan a safe path, which may lead people into a dangerous area. That is, the existing system often lacks comprehensive perception ability of real-time environmental factors, which makes it difficult to consider the influence of multi-dimensional factors such as temperature, smoke concentration and personnel density on evacuation safety. Therefore, a reinforcement learning driven fire emergency evacuation guidance system is proposed to solve this problem. SUMMARY

[0005] To solve the above technical problems, a reinforcement learning driven fire emergency evacuation guidance system is provided, which solves the problems proposed in the background art.

[0006] In order to achieve the above purpose, the technical scheme of the present application is as follows:

[0007] The present application provides a reinforcement learning driven fire emergency evacuation guidance system, characterized in that it comprises:

[0008] A data acquisition module for real-time acquisition of environmental data of a plurality of predetermined positions in the building and a plurality of standard evacuation paths, wherein the environmental data includes temperature, smoke concentration and personnel density;

[0009] A hazard value acquisition module for dividing the predetermined positions into a plurality of grid points according to a preset size, normalizing and weightedly fusing the environmental data in each grid point to obtain a comprehensive hazard value;

[0010] a candidate path acquisition module configured to eliminate grid points with a comprehensive risk value greater than a preset risk threshold value from the grid points, and input the comprehensive risk values of the remaining grid points after the elimination into an enhanced learning decision model to generate a complete candidate path sequence;

[0011] a safety score calculation module configured to calculate an initial safety score of the node sequence of the candidate path sequence by matching the node sequence with a node sequence of a standard evacuation path, calculate a risk difference value between the comprehensive risk value of each grid point in the path and a preset risk threshold value and a density difference value between a personnel density normalization value and a preset density threshold value, and adjust the initial safety score by deducting points according to the risk difference value and the density difference value to obtain a path safety score;

[0012] an efficiency score calculation module configured to calculate an initial efficiency score by calculating a ratio of a total length of the candidate path sequence to a preset standard path length, and adjust the initial efficiency score by deducting points according to the risk difference value and the density difference value to obtain a path efficiency score;

[0013] an optimal path acquisition module configured to select an optimal evacuation path from the candidate path sequence based on the path safety score and the path efficiency score by a multi-objective optimization algorithm, and send data of the optimal evacuation path to an intelligent indication device in the building for path guidance.

[0014] Compared with the prior art, the beneficial effects of the present application are:

[0015] The present application generates a standard evacuation path based on building layout and safety specifications, generates a candidate path sequence by eliminating high-risk grid points and inputting the comprehensive risk values of the remaining grid points into an enhanced learning decision model, solves the problem that a fixed evacuation scheme cannot adapt to a dynamic fire environment, and realizes intelligent path planning based on real-time risk distribution.

[0016] The present application obtains an initial safety score by calculating the matching degree of the candidate path and the standard path, dynamically adjusts the score by combining the difference between the risk values of the grid points in the path and the threshold value and according to the personnel density, obtains an initial efficiency score based on the path length ratio, and optimizes the adjustment based on the difference between the risk values and the threshold value and the personnel density, solves the technical problem that the traditional system is difficult to balance between safety and efficiency, and realizes scientific decision-making of multi-objective collaborative optimization. BRIEF DESCRIPTION OF DRAWINGS

[0017] The disclosure of the present application will be described with reference to the accompanying drawings. It should be understood that the drawings are only for illustrative purposes, and are not intended to limit the scope of protection of the present application. Among them:

[0018] Figure 1 a structure diagram of the enhanced learning driven fire emergency evacuation guidance system proposed in the present application;

[0019] Figure 2 Method flow chart implemented for the system in the present application;

[0020] Figure 3 Data flow chart implemented for the system in the present application;

[0021] Figure 4 Method flow chart of dynamic optimization of candidate path sequence proposed in the present application. DETAILED DESCRIPTION

[0022] It is easy to understand that, according to the technical solution of the present application, those skilled in the art can propose a plurality of structure modes and implementation modes which can be replaced with each other without changing the essential spirit of the present application. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical solution of the present application, and should not be regarded as the whole or as the limitation or restriction of the technical solution of the present application.

[0023] In the prior art, the fire emergency evacuation system has the problems of static solidification of path planning, insufficient environment situation awareness, and difficulty in balancing safety and efficiency. It usually adopts a guiding mode of preset fixed path. Once the original path becomes impassable due to changes in fire situation, the system often cannot autonomously perform dynamic re-planning and optimization adjustment, but still relies on the pre-set evacuation scheme. This static and passive nature reduces the evacuation efficiency.

[0024] In order to solve the above problems, the present application collects environmental data in real time and divides them into grid points, normalizes and weights multi-dimensional environmental factors to integrate them into a comprehensive risk value, generates standard paths based on building layout and eliminates high-risk grid points, inputs the risk values of the remaining grid points into an enhanced learning model to generate a candidate path sequence. Through dual evaluation of the candidate path in terms of safety score and efficiency score, the safety score is calculated by path sequence matching degree and dynamically adjusted in combination with risk value difference, and the efficiency score is calculated by path length ratio and optimized in combination with personnel density influence. Finally, the optimal evacuation path is selected based on a multi-objective optimization algorithm and sent to an intelligent indication device. Through dynamic path planning driven by enhanced learning and multi-objective optimization decision, the present application realizes real-time optimization and adaptive adjustment of evacuation paths, solves the problem of poor adaptability of static guiding systems in complex fire environment, ensures safety while improving evacuation efficiency, and enhances the reliability and effectiveness of the emergency evacuation system.

[0025] As shown in Figure 1 Each module of the enhanced learning driven fire emergency evacuation guiding system is introduced, including:

[0026] The data acquisition module 100 is used to collect environmental data of a plurality of predetermined points in a building and a plurality of standard evacuation paths in real time. The environmental data includes temperature, smoke concentration and personnel density.

[0027] a hazard value obtaining module 200, configured to divide the predetermined position points into a plurality of grid points according to preset sizes, normalize and perform weighted fusion processing on the environmental data in each grid point to obtain comprehensive hazard values;

[0028] a candidate path obtaining module 300, configured to eliminate the grid points with the comprehensive hazard values greater than a preset hazard threshold value, and input the comprehensive hazard values of the remaining grid points after the elimination into an enhanced learning decision model to generate a complete candidate path sequence;

[0029] a safety score calculating module 400, configured to calculate a matching degree between a node sequence of the candidate path sequence and a node sequence of a standard evacuation path to obtain an initial safety score, calculate a hazard difference value between the comprehensive hazard value of each grid point in the path and a preset hazard threshold value and a density difference value between a personnel density normalized value and a preset density threshold value, and adjust the initial safety score according to the hazard difference value and the density difference value to obtain a path safety score;

[0030] an efficiency score calculating module 500, configured to calculate a ratio between a total length of the candidate path sequence and a preset standard path length to obtain an initial efficiency score, and adjust the initial efficiency score according to the hazard difference value and the density difference value to obtain a path efficiency score;

[0031] an optimal path obtaining module 600, configured to select an optimal evacuation path from the candidate path sequence based on the path safety score and the path efficiency score through a multi-objective optimization algorithm, and send data of the optimal evacuation path to an intelligent indication device in the building to guide the path.

[0032] As shown in Figure 2 , the specific implementation manners of the modules in the enhanced learning driven fire emergency evacuation guidance system are introduced.

[0033] Regarding the data obtaining module 100:

[0034] The environmental data is collected in real time by a multi-source sensor network deployed in the building, specifically including:

[0035] The personnel density is obtained by using a target detection algorithm based on deep learning to count the number of personnel in a unit area in real time through visual sensors and infrared thermal imagers deployed at entrances and exits of the channels and rooms.

[0036] The smoke concentration is detected by smoke sensors distributed on the top of the building, and the temperature in the vertical direction is collected by temperature sensor arrays installed on the walls.

[0037] The timestamp of the data collection is obtained, and the above data is synchronized based on the timestamp, and the spatial coordinate data of each sensor is obtained from the building information model.

[0038] Quality optimization and feature extraction are performed on the collected raw environmental data, specifically including:

[0039] Sensor abnormal readings are identified using a sliding window detection method, and abnormal data is repaired based on the spatial correlation of adjacent sensor data, specifically including: based on spatio-temporal correlation analysis, using the environmental data of adjacent grid points to complete the missing data through spatio-temporal Kriging interpolation algorithm, and identifying and correcting abnormal data through data consistency test; sensor noise is eliminated through Kalman filtering algorithm to improve data stability;

[0040] Statistical features, including mean, variance, peak value and gradient, are extracted from the raw sensor data, and the dimension and value of each sensor data are unified and normalized to eliminate data bias caused by sensor differences;

[0041] Standardized environmental data is generated, specifically including:

[0042] The environmental data is organized according to the predetermined site number, and each data record contains a timestamp, spatial coordinates, personnel density, temperature, and smoke concentration;

[0043] The process of obtaining a standard evacuation path includes:

[0044] A digital floor plan or building information data of the building is obtained, including room distribution, corridor structure, stair location, elevator room, safe exit location, and passage width size;

[0045] and the evacuation path design specification is obtained, including the maximum evacuation distance requirement, the minimum number of safety exits, the width standard of escape passage, and the obstacle avoidance rule;

[0046] Based on the obtained data, a standard evacuation path that meets the specification is generated, specifically including:

[0047] The building layout data is parsed, and all safety exits, stairwells, corridor intersections, and large space center points are identified as key nodes of the path, and a node position coordinate database is established;

[0048] Based on the connectivity between nodes, all feasible path segments are generated, and the geometric length, difficulty coefficient and potential capacity limit of each path segment are calculated;

[0049] The key nodes are taken as vertices and the path segments are taken as edges to construct an evacuation path graph, and a graph traversal algorithm is applied to generate multiple candidate paths from each starting point to the safety exit, including the shortest path, backup path and redundant path, to ensure path diversity;

[0050] Compare the generated paths with the evacuation path design specifications to verify whether each path meets the evacuation distance restrictions, exit capacity requirements, and passage width standards. Modify or remove paths that do not meet the specifications.

[0051] Generate standardized format evacuation path data, including:

[0052] Structured path storage: store all standard evacuation paths in a graph data structure. Each path record contains a node sequence, a path segment list, and the total length of the path.

[0053] Path attribute labeling: label each path with attribute information, including path type, priority level, applicable scenario, and expected travel time.

[0054] Encapsulate the above data into a unified data packet and transmit it to the central processing system through industrial Ethernet. Use CRC (Cyclic Redundancy Check) verification mechanism to ensure data integrity during transmission.

[0055] Provide data query services to other system modules through standard API interfaces, supporting real-time data stream access and historical data retrieval.

[0056] Regarding the hazard value acquisition module 200:

[0057] Acquire real-time environmental data of each predetermined location provided by the data acquisition module, and read preset grid size parameters from the configuration file, including grid length, width, and height information, grid overlap rate settings, and pre-set initial weight coefficients of each environmental factor, as well as weight dynamic adjustment rules.

[0058] Grid processing and hazard value calculation of acquired data, including:

[0059] Divide the building space into regular grid units based on the preset size, establish a grid coordinate system, and map each predetermined location to the corresponding grid unit.

[0060] Standardize the environmental data in each grid unit and process it using the min-max normalization method.

[0061] Adjust the weight coefficients of each factor according to the real-time change rate of environmental data, and assign higher weights to rapidly growing data.

[0062] Weighted sum of normalized environmental factors according to adjusted weight coefficients to calculate the comprehensive hazard value of each grid unit.

[0063] Generate standardized comprehensive hazard value data, including:

[0064] The comprehensive risk value data is organized by grid cell number, and each grid cell data record contains grid coordinates, a comprehensive risk value, normalized values of each environmental data, and a timestamp;

[0065] Specifically, the process of calculating the comprehensive risk value includes:

[0066] The following data is obtained:

[0067] Environmental base data: real-time environmental monitoring data of each predetermined site is obtained from the data acquisition module, including temperature measurement value , smoke concentration detection value , and personnel density statistical value , wherein represents the number of predetermined sites;

[0068] Grid division parameters: read the preset grid size specifications from the system configuration library, including grid cell length , and width , and grid overlap coefficient ;

[0069] Weight configuration parameters: read the basic weight distribution of each environmental factor from the expert knowledge base , , , and weight dynamic adjustment parameters.

[0070] Based on the obtained data, the following calculation steps are performed:

[0071] (1) Planar grid division

[0072] Based on the preset size parameters, the three-dimensional space of the building is divided into a regular grid cell system:

[0073] ;

[0074] ;

[0075] wherein represents the total length of the building, represents the total width of the building, represents the number of grid cells divided in the x direction, represents the number of grid cells divided in the y direction, represents the upward rounding operation.

[0076] A grid coordinate system is established , wherein , respectively represent the coordinate index of the grid in the two-dimensional plane. Each predetermined site is associated to the corresponding grid cell through a mapping function .

[0077] (2) Data standardization processing

[0078] The environmental monitoring data in each grid cell is standardized and converted, and the min-max normalization method is used to obtain , , ; wherein , , are the normalized values of temperature, smoke concentration and personnel density, with a value range of [0, 1].

[0079] (3) Dynamic adjustment of weights

[0080] According to the real-time change trend of environmental data, the weight distribution of each factor is dynamically adjusted, and higher weight coefficients are given to rapidly growing risk factors:

[0081] ;

[0082] ;

[0083] ;

[0084] wherein: , , represent the real-time change rate of temperature, smoke concentration and personnel density, is the change sensitivity coefficient, which controls the influence degree of change rate on weight distribution.

[0085] The adjusted weights are normalized:

[0086] ;

[0087] ;

[0088] ;

[0089] (4) Calculation of comprehensive risk value

[0090] The normalized environmental factors are weighted and fused according to the adjusted weight coefficients, and the comprehensive risk value of each grid cell is calculated:

[0091] ;

[0092] wherein represents the comprehensive risk value of the grid cell , with a value range of [0, 1], and the larger the value, the higher the risk degree of the grid cell.

[0093] Generating standardized output data:

[0094] Integrated risk value data structure: Organize integrated risk value data by grid cell number, each grid cell data record contains: grid coordinates, integrated risk value, normalized values of each environmental data and data calculation timestamp.

[0095] Regarding the candidate path acquisition module 300:

[0096] Acquire the integrated risk value of each grid point and its spatial distribution data, read the preset risk threshold parameter (such as set to 0.8) from the system configuration library, and load the pre-trained reinforcement learning decision model, including the policy network parameter.

[0097] Based on the acquired data, grid screening and path sequence generation are performed, specifically including:

[0098] Traverse the integrated risk value of all grid points, and mark the grid points with integrated risk value exceeding the risk threshold as impassable areas;

[0099] Based on the elimination result, the feasible passable area is reconstructed, and the connectivity of the grid points adjacent to the dangerous area is detected to ensure the integrity of the feasible area;

[0100] Encode the integrated risk value, passable state mark and spatial position information of the remaining grid points into a state representation that can be processed by the reinforcement learning model;

[0101] Input the state data into the reinforcement learning decision model, calculate the selection probability of each actionable action through the policy network, and select the optimal action sequence based on the probability distribution;

[0102] From the starting position, gradually expand the path nodes according to the action sequence output by the reinforcement learning model until reaching the safe exit, forming a complete candidate path sequence;

[0103] Physically check the generated candidate path sequence, including path connectivity verification and channel width adaptability evaluation;

[0104] Generating standardized candidate path sequences, specifically including:

[0105] Organize each candidate path sequence into an ordered node list, each node containing spatial coordinates, grid risk value and passable state information;

[0106] Further, based on the predicted updated integrated risk value, the candidate path sequence is optimized, specifically including:

[0107] Continuously acquire historical time series data of grid point environmental data within the calibration time window;

[0108] The prediction value is obtained by analyzing and predicting the historical time series data based on a time series analysis algorithm, when the prediction value is greater than a danger threshold, marking the grid point as a dangerous diffusion area, and replacing the comprehensive danger value of the grid point with the prediction value, and eliminating the candidate path containing the grid point from the candidate path sequence and updating the candidate path sequence.

[0109] Exemplarily, during the evacuation process, the historical environmental data of each grid point is continuously acquired to form the historical time series data. For a certain grid point, the system uses a time series analysis algorithm (such as exponential smoothing method) to predict the trend of these data. Specifically, the algorithm fits a curve based on historical data points to predict the comprehensive danger value at a future time (such as 30 seconds in the future), and when the predicted comprehensive danger value is greater than a danger threshold (such as set to 0.9), the prediction value is taken as the updated comprehensive danger value of the grid point.

[0110] Through the above technical solutions, the application realizes the prediction and update of the comprehensive danger value of the grid point, and provides a decision basis for the real-time optimization of the evacuation path. By using a time series analysis algorithm to predict the trend of the environmental data, the system can identify the dangerous diffusion area in advance, and mark and eliminate the related grid points in time when the danger threshold is exceeded, thereby effectively avoiding the entry of personnel into the environment that is about to deteriorate. This mechanism not only improves the safety and reliability of path planning, but also enhances the adaptive ability of the system to sudden environmental changes, ensuring that the evacuation process is always based on the latest risk situation for dynamic adjustment. In addition, by replacing the real-time data with the prediction value, the system retains the responsiveness while reducing the risk of misjudgment caused by lagging environmental data, further optimizing the evacuation efficiency.

[0111] Further, the construction process of the reinforcement learning decision model specifically includes:

[0112] acquiring historical environmental data and path selection decisions of the grid points of the building;

[0113] abstracting the historical environmental data into a discrete state set to construct a state space;

[0114] abstracting the path selection decisions into a discrete action set to construct an action space;

[0115] constructing a reward function based on a safety reward and an efficiency reward, wherein the safety reward is calculated based on the difference between the comprehensive danger value of the path and the danger threshold, and the efficiency reward is determined based on the ratio of the path length to the standard path length;

[0116] constructing a policy network by fitting a probability distribution mapping relationship from the state space to the action space through a deep neural network;

[0117] constructing a decision model based on the reward function and the policy network.

[0118] Specifically, in the building, first obtain the historical environmental data of the building grid points, including but not limited to temperature, smoke concentration and personnel density. The historical environmental data is abstracted into a discrete state set, thereby constructing a state space. Specifically, the environmental data of each grid point is normalized and divided into multiple danger levels, for example, the smoke concentration is divided into "safe", "warning" and "danger" three levels based on the preset threshold. Each state in the state space represents the joint combination of the danger levels of all grid points in the building, thereby capturing the overall environmental condition.

[0119] The path selection decision is abstracted into a discrete action set, thereby constructing an action space. The action space includes all possible directions of moving from the current grid point to the adjacent grid point, such as "up", "down", "left", "right", etc. Each action corresponds to a moving step in the evacuation path, and the action space covers all reachable moving options in the building.

[0120] The reward function is constructed based on the safety reward and the efficiency reward. The safety reward is calculated according to the difference between the comprehensive danger value of the path and the safety threshold: when the comprehensive danger value (calculated by weighted average) of all grid points on the path is lower than the safety threshold, a positive reward is given; otherwise, a negative reward is given. The efficiency reward is determined based on the ratio of the path length to the standard path (i.e. the preset shortest path length in the building): the shorter the path length, the higher the efficiency reward; for example, if the ratio of the path length to the standard path is less than 1, a positive reward is given, otherwise a negative reward is given. The reward function combines the safety reward and the efficiency reward by weighted combination to balance the evacuation safety and efficiency.

[0121] The policy network is implemented by a deep neural network that fits the probability distribution mapping relationship from the state space to the action space. The deep neural network adopts a multi-layer perceptron structure, the input layer receives the state vector (representing the current environmental state), and the output layer generates the probability value of each possible action. The network learns and optimizes the weight parameters through training, so that given a state, the output action probability tends to be a high reward decision.

[0122] Finally, a decision model is constructed based on the reward function and the policy network. The model is trained using reinforcement learning algorithms (such as policy gradient method), and the network parameters are updated iteratively to maximize the cumulative reward. During the training process, the model simulates multiple evacuation scenarios and continuously adjusts the strategy to generate safe and efficient evacuation paths. The decision model is integrated into the intelligent evacuation system for real-time path planning and dynamic adjustment.

[0123] Through the technical solution, the application realizes the intelligent evacuation decision model construction based on reinforcement learning, and provides a data-driven decision basis for dynamic path planning. By abstracting the historical environment data and path selection decision as state space and action space respectively, a calculable mapping relationship between the environment and action is established. Based on the design of a composite reward function of safety reward and efficiency reward, an effective balance between danger avoidance and evacuation efficiency of path selection is realized, wherein the safety reward quantifies the risk by the difference between the comprehensive danger value and the safety threshold, and the efficiency reward optimizes the time cost by the ratio of path length to standard path. By fitting the probability distribution of state space to action space through a deep neural network, the policy network can adaptively learn the optimal strategy, improving the decision accuracy and response speed of the model in complex environments. In addition, the structured processing and normalization method of historical data enhances the compatibility of the model to multi-source data, enabling the system to have continuous learning and optimization capabilities, further improving the adaptability and reliability of evacuation planning.

[0124] Further, as shown in Figure 3 , a process of dynamically optimizing the candidate path sequence is introduced, specifically including:

[0125] obtaining the generated candidate path sequence and calculating the path continuity in the sequence;

[0126] If the path continuity is less than the continuity threshold, the candidate path is discarded;

[0127] Calculate the minimum personnel density that the path node in the remaining candidate path can accommodate. If the minimum personnel density is less than the current personnel density, the candidate path is discarded, and the remaining candidate path is taken as the optimized candidate path sequence.

[0128] Further, the diversity of the dynamically optimized candidate path sequence is guaranteed, specifically including:

[0129] Calculate the path repetition degree of the path coverage area in the optimized candidate path sequence;

[0130] When the path repetition degree exceeds the preset repetition degree threshold, generate a supplementary path based on the candidate path sequence before optimization.

[0131] Illustratively, after generating the initial candidate path sequence, the following optimization process is performed:

[0132] (1) Dynamic optimization of candidate path sequence

[0133] First, obtain the candidate path sequence generated by the reinforcement learning model. The system evaluates the continuity of each candidate path in the sequence, specifically including:

[0134] Path continuity calculation: The system analyzes the topology of each path and detects whether there are unreasonable turns or U-turns. For example, the system calculates the angle of direction change between adjacent corridor segments in the path. If the number of consecutive turns with an angle of 90 degrees exceeds a preset value (e.g., 3 times), the system determines that the path continuity is poor. The system sets a continuity threshold (e.g., 0.7, based on the path smoothness index). If the calculated path continuity is less than the threshold, the system removes the candidate path from the sequence. For example, a path that requires a person to make three right-angle turns within a 20-meter distance is automatically excluded by the system because it does not conform to the natural behavior pattern of personnel evacuation.

[0135] Node capacity evaluation: For the remaining candidate paths that pass the continuity test, the system further calculates the minimum personnel density that each key node (e.g., narrow corridor, security gate, stair platform) can accommodate. This minimum personnel density is based on building design specifications and historical evacuation data. The system compares the minimum personnel density of the node with the current personnel density obtained in real time. If the minimum personnel density of any node is less than the current personnel density (e.g., a certain stair platform has a minimum capacity of 3 people per square meter, while the current monitoring density has reached 4 people per square meter), the system determines that the path has a congestion risk and removes it from the candidate sequence.

[0136] Through the above screening, the optimized candidate path sequence is finally obtained, ensuring that all paths have reasonable continuity and the capacity of each node meets the current evacuation demand.

[0137] (2) Diversity guarantee of candidate path sequence

[0138] To prevent the optimized candidate paths from being overly concentrated in certain areas, the system performs the following diversity guarantee measures:

[0139] Path repetition calculation: The system calculates the coverage area of all paths in the optimized candidate path sequence on the building plan and analyzes the degree of sharing of corridor segments by different paths. Specifically, the system quantifies the path repetition by calculating the Jaccard similarity coefficient of the path set. When the repetition exceeds a preset threshold (e.g., 0.6), it indicates that the path diversity is insufficient.

[0140] Supplementary path generation: When the path repetition is detected to be excessive, the system reverts to the candidate path sequence before optimization and selects paths that meet the following conditions from the excluded paths for supplementation:

[0141] Paths that are slightly below the threshold in terms of continuity but have a topology that is significantly different from other paths;

[0142] Paths that fail the node capacity evaluation but show that the personnel density has decreased to a safe level through real-time monitoring.

[0143] By the technical solution, the multi-dimensional optimization and diversity guarantee of the candidate path sequence are achieved, and the rationality and reliability of the evacuation path planning are improved. Through the dual screening mechanism of path coherence evaluation and node capacity detection, paths with unreasonable turning or potential congestion risk are effectively eliminated, ensuring the feasibility and safety of the candidate paths. Meanwhile, through the path repetition degree analysis and supplementary path generation mechanism, the excessive concentration of evacuation paths in spatial distribution is avoided, ensuring the reasonable coverage of the path network and the resource utilization efficiency. This combination of optimization and guarantee maintains the agility of path planning in responding to real-time environmental changes, and enhances the resilience of the system in dealing with unexpected situations through diversity maintenance, thereby providing a more comprehensive and reliable path solution for personnel evacuation.

[0144] Regarding the safety score calculation module 400:

[0145] The safety score calculation process is as follows:

[0146] The following data is obtained:

[0147] Candidate path sequence data: a set of generated candidate paths obtained from the candidate path generation module Each candidate path contains a node sequence , where is the number of nodes of the path , and each node contains coordinate information and a path identifier.

[0148] Standard evacuation path data: a set of pre-stored standard evacuation paths obtained from the standard path acquisition module Each standard path contains a node sequence , where is the number of nodes of the standard path .

[0149] Grid danger value data: the comprehensive danger value of each grid point obtained from the danger value acquisition module. Let the grid point set be , and the comprehensive danger value of each grid point is , and the mapping relationship data between the grid point and the path node is provided, i.e. for the node of the path , its corresponding grid point is .

[0150] Personnel density distribution data: personnel density value, the normalized value of the personnel density of each grid point is .

[0151] Threshold parameter data: Read preset danger thresholds from the system configuration library. (e.g., set to 0.9) and preset density threshold (For example, set it to 0.9).

[0152] Based on the acquired data, the following steps are performed to calculate and adjust the security score:

[0153] (1) Calculation of path sequence matching degree

[0154] Candidate paths Node sequence Standard evacuation routes Node sequence Node-by-node comparison is performed, and a similarity algorithm based on edit distance is used to calculate the sequence matching degree. Edit distance Defined as a sequence Convert to Minimum number of editing operations (insert, delete, replace) required. Matching degree. The calculation is as follows:

[0155] ;

[0156] in, Representing a path Compared to the standard path Edit distance between Representing a path The length of the node sequence, Indicates standard path The length of the node sequence.

[0157] Initial security score Based on optimal matching degree allocation:

[0158] ;

[0159] The initial security score ranges from 0 to 100, with higher values ​​indicating a closer match to the standard path.

[0160] (2) Risk value difference analysis

[0161] Traverse candidate paths Each node Obtain its corresponding grid points Overall risk value Calculate the danger difference for each node. :

[0162] ;

[0163] Establish a risk level classification table, based on The magnitude of the hazard is categorized into several levels (e.g., mild, moderate, severe). Let the categorization threshold be... and (For example , ),but:

[0164] like The risk level is mild.

[0165] like The risk level is moderate.

[0166] like The danger level is severe.

[0167] (3) Safety score deduction adjustment

[0168] Extraction path Normalized values ​​of personnel density at grid points corresponding to each node .

[0169] Based on the hazard level classification table and differences in personnel density, the initial safety score is dynamically adjusted by deductions. Deductions include hazard deductions and density deductions.

[0170] Dangerous points deduction Points are deducted from each node based on its hazard level. Let the deduction coefficient for a minor hazard be... The moderate deduction coefficient is The severe deduction coefficient is (For example , , ).but:

[0171] ;

[0172] in according to The value of the corresponding level, This is an indicator function (value 1 when the condition is true, otherwise 0).

[0173] Density deduction Points are deducted based on differences in personnel density. The density difference at each node is calculated. And set the density deduction factor as (For example ).but:

[0174] ;

[0175] Adjusted security score The calculation is as follows:

[0176] ;

[0177] (4) Safety score normalization

[0178] Map the adjusted safety scores to the normalized score interval [0, 100] to ensure the comparability of safety scores of different paths. The final safety score is calculated as follows:

[0179] .

[0180] Regarding the efficiency score calculation module 500:

[0181] The following data is obtained:

[0182] Candidate path sequence data: Obtain the complete candidate path sequence from the candidate path acquisition module, including the node coordinate sequence and path segment length data of each path;

[0183] Standard path length data: Obtain the reference length value of the preset standard path and its corresponding node sequence information from the standard path acquisition module;

[0184] Based on the obtained data, efficiency score calculation and optimization adjustment are performed, specifically including:

[0185] Based on the node coordinate data of the candidate path sequence, the total length of each candidate path is accurately calculated using geometric algorithms;

[0186] Compare and analyze the total length of the candidate path with the preset standard path length, calculate the length ratio, and assign an initial efficiency score based on the length ratio;

[0187] Iterate through each grid point in the candidate path sequence, calculate the absolute difference between the comprehensive risk value of each grid point and the preset risk threshold, and establish a risk level table;

[0188] Extract the personnel density normalized value of the grid point corresponding to each node of the path, calculate the overall personnel density characteristic value and density distribution uniformity index of the path;

[0189] Generate standardized path efficiency score data, specifically including:

[0190] Efficiency score data structuring: generate complete efficiency score data records for each candidate path, including the final efficiency score value, initial efficiency score value, deduction details, and efficiency level label.

[0191] Specifically, the process of calculating the efficiency score includes:

[0192] Obtain candidate path sequence data: obtain the complete candidate path sequence set from the candidate path acquisition module Each candidate path contains a node coordinate sequence , wherein represents the node coordinates, is the path number of nodes, while containing the length data of each path segment , wherein represents the node to segment length.

[0193] And the standard path length data: the standard path acquisition module acquires the reference length value of the preset standard path and its corresponding node sequence information, the standard path length is the average of the lengths of all standard evacuation paths.

[0194] Based on the acquired data, the following steps are performed for efficiency score calculation and optimization adjustment:

[0195] (1) Candidate path length accurate calculation

[0196] Based on the node coordinate sequence of the candidate path , the total length of the path is accurately calculated using a geometric algorithm :

[0197] ;

[0198] (2) Initial efficiency score calculation

[0199] Compare and analyze the total length of the candidate path with the preset standard path length , calculate the initial efficiency score :

[0200] ;

[0201] (3) Efficiency score deduction adjustment

[0202] The adjusted efficiency score is calculated as follows:

[0203] ;

[0204] (4) Efficiency score normalization processing

[0205] Map the adjusted efficiency score to the standardized score interval [0, 100] to ensure the comparability of the efficiency scores of different paths. The final safety score is calculated as follows:

[0206] .

[0207] Regarding the optimal path acquisition module 600:

[0208] ​Obtain path security score, path efficiency score, and candidate path sequence;

[0209] And preset weight parameters for multi-objective optimization, including safety weights, efficiency weights, and dynamic adjustment rules;

[0210] And obtain online status and working status information of each indicator device from the intelligent indicator device management system.

[0211] Optimal path decision-making and verification are performed based on the acquired data, specifically including:

[0212] Establish a weighted comprehensive evaluation function that combines safety and efficiency scores, dynamically adjust the weight ratio of safety and efficiency according to the current emergency level, and calculate the comprehensive score of each candidate path.

[0213] Specifically, the process of calculating the overall score includes:

[0214] (1) Establish a comprehensive evaluation function

[0215] Construct a weighted comprehensive evaluation function that combines safety score and efficiency score:

[0216] ;

[0217] in, Representing a path The overall score, Representing a path Security score, Representing a path The efficiency score, This represents the dynamic weight of the security score. The dynamic weights representing the efficiency scores satisfy the following constraints: ;

[0218] (2) Dynamic adjustment of weights

[0219] According to the current emergency level Dynamically adjust weight allocation. Set emergency levels. The value range is [0, 1], where 0 represents the normal state and 1 represents the highest emergency state. For example, mild cases take a value of 0.3, moderate cases take a value of 0.6, and moderate cases take a value of 0.9.

[0220] The weight adjustment uses a linear interpolation method:

[0221] ;

[0222] ;

[0223] in, This represents the base security weight (usually set to a high value, such as 0.7). represents the base efficiency weight (usually set to a low value, such as 0.3), represents the safety weight decay coefficient, which controls the degree of influence of the emergency level on the safety weight, represents the efficiency weight enhancement coefficient, which controls the degree of influence of the emergency level on the efficiency weight;

[0224] normalize the calculated weights to ensure :

[0225] ;

[0226] (3) Path feasibility verification

[0227] Combine real-time environmental monitoring data to verify the feasibility of the candidate path:

[0228] ;

[0229] The feasibility judgment is based on the comparison of real-time monitoring values of environmental parameters such as temperature and smoke concentration with preset critical thresholds.

[0230] (4) Comprehensive score calculation

[0231] Combine the feasibility verification results to calculate the final comprehensive score of each candidate path:

[0232] ;

[0233] For infeasible paths ( =0), the comprehensive score is directly set to 0 to ensure that it will not be selected as the optimal path.

[0234] Use multi-objective optimization algorithm to identify non-dominated solution set, and select path subset that achieves Pareto optimality in safety and efficiency from candidate paths;

[0235] Combine the latest environmental monitoring data to perform secondary verification on the optimized path to ensure that the risk values of each grid point in the path have not deteriorated;

[0236] Analyze the node sequence connectivity of the optimized path to identify possible breakpoints or obstacles and ensure the executability of the path in the actual environment.

[0237] Generate the final optimal path decision and execute the guidance instructions, which specifically include:

[0238] Generate an optimal evacuation path containing path node sequence, guidance strategy, and estimated travel time, etc. Complete information;

[0239] Convert the optimal path data into control instruction sequences that can be recognized by intelligent indication devices, including display content and indication direction;

[0240] The guiding instructions are distributed to the relevant intelligent indication devices through the Internet of Things communication protocol, ensuring accurate delivery and execution of the instructions.

[0241] Further, the effectiveness of the optimal evacuation path is continuously monitored, specifically including:

[0242] Real-time acquisition of environmental monitoring data of the optimal evacuation path, and prediction of the remaining safety time of the path based on the change trend of the environmental monitoring data;

[0243] When the remaining safety time is less than a preset time threshold, the optimal evacuation path is re-planned.

[0244] Further, the cooperative guiding mechanism of the intelligent indication device specifically includes:

[0245] Acquisition of the layout position and state information of each intelligent indication device in the building;

[0246] Based on the optimal evacuation path and the layout position of the device, a hierarchical guiding instruction sequence is generated, including a main guiding scheme and a backup guiding scheme;

[0247] Monitoring the state information of the device, when the device is normal, the main guiding scheme is enabled, and when a device failure is detected, the backup guiding scheme is enabled.

[0248] Further, it also includes real-time monitoring of the evacuation effect, specifically including:

[0249] Acquisition of the personnel density in the actual evacuation process;

[0250] Calculation of the density difference between the personnel density and the density threshold, and adjustment of the strategy network of the model according to the density difference;

[0251] Replanning of the candidate path sequence according to the adjusted model, updating of the optimal evacuation path according to the re-planned candidate path sequence, and real-time updating of the path guidance of the intelligent indication device.

[0252] For example, in a building, an initial optimal evacuation path is acquired.

[0253] (1) Continuous monitoring of the effectiveness of the optimal evacuation path

[0254] Real-time environmental data, such as temperature, smoke concentration, and personnel density, are collected along the optimal evacuation route by environmental monitoring sensor networks. Based on these data, the system uses time series analysis methods (such as exponential smoothing or ARIMA models) to predict the trend of environmental deterioration, thereby calculating the remaining safety time of the route. The remaining safety time is defined as the estimated time for the route to remain passable. The system has a preset time threshold (e.g., 3 minutes), and when the prediction shows that the remaining safety time is less than this threshold, it immediately considers real-time environmental data, building structure constraints, and personnel distribution to generate a new candidate path sequence and update the optimal evacuation route. For example, if the original path leads to a sharp rise in smoke concentration due to the spread of fire, the system will quickly calculate an alternative path to avoid personnel entering a dangerous area.

[0255] (2) Collaborative guidance mechanism of intelligent indication devices

[0256] The layout positions (such as installation at corridor intersections, stair entrances) and state information (such as normal operation, failure, or low power) of all intelligent indication devices in the building are obtained. Based on the current optimal evacuation route, the system generates a hierarchical guidance instruction sequence, where the main guidance scheme specifies that the indication devices along the optimal path display green arrows and voice prompts, and the backup guidance scheme specifies that alternative devices provide detour guidance when the devices fail. The system monitors the status of each device in real time, and when it detects that a certain electronic sign is malfunctioning, it automatically enables the backup guidance scheme, such as issuing auditory guidance through sound and light alarms near the failure point. In addition, the system dynamically adjusts the guidance scheme based on data feedback from personnel density detection devices: if the personnel density on the main path exceeds a preset threshold (e.g., 2 people per square meter), the system enables the backup guidance scheme to divert personnel, such as guiding some personnel to evacuate through secondary channels, to balance the load.

[0257] (3) Real-time monitoring mechanism of evacuation effect

[0258] During the evacuation process, personnel density detection devices continuously collect actual personnel density data. The density difference between the real-time personnel density and the density threshold (based on the channel capacity) is calculated, and the policy network of the reinforcement learning model is adjusted online based on this difference. Specifically, the network weights are updated through the backpropagation algorithm to optimize the path selection strategy, such as preferentially selecting wide or low-traffic paths when the density difference is too large. Based on the adjusted model, the candidate path sequence is re-planned, the optimal evacuation route is updated, and the guidance information of the intelligent indication devices is updated synchronously. For example, if the personnel gather near the main exit, causing a sharp increase in density, the system will recalculate the path to guide subsequent personnel to evacuate through the backup exit, while dynamically adjusting the display content of the indication devices to ensure that the guidance information is consistent with the latest path.

[0259] By the technical solution, the application constructs a complete evacuation management system integrating path dynamic monitoring, intelligent device cooperation and evacuation effect feedback. Through residual safety time prediction and threshold triggering mechanism, the optimal evacuation path is maintained and updated in advance, effectively avoiding path failure risk. Through the hierarchical guidance scheme and the adaptive cooperation mechanism of device state, the robustness of the guidance system is ensured, and dynamic load balancing based on personnel density is realized. Further, the strategy network is adjusted online through personnel density feedback, forming a closed-loop optimization of perception-decision-execution, so that the system has continuous self-improvement ability. This multi-level and adaptive integrated design not only ensures the safety of evacuation, but also improves the overall evacuation efficiency and system reliability.

[0260] As shown in Figure 4 , a data flowchart of the system implementation is introduced.

[0261] In order to facilitate the understanding of the above embodiments, the following will take one specific application scenario of the above embodiments as an example for illustration:

[0262] Scenario: Fire emergency evacuation guidance in multi-story office building

[0263] (1) Data acquisition module

[0264] Multi-source sensor networks are deployed in the corridors and rooms of the three-story office building, including visual sensors, infrared thermal imagers, smoke sensors and temperature sensors.

[0265] Specific operation:

[0266] Real-time collection of environmental data at predetermined points (such as corridor intersections and room entrances). For example:

[0267] Point A: Personnel density is 0.8 person / m2 (detected by visual sensor and infrared thermal imager).

[0268] Point B: Smoke concentration is 200 ppm (detected by smoke sensor).

[0269] Point C: Temperature is 50°C (detected by temperature sensor array).

[0270] After cleaning and fusion, the standardized environmental data set is generated.

[0271] Output structured environmental data, including timestamp, spatial coordinates, personnel density, smoke concentration and temperature value.

[0272] Standard path acquisition:

[0273] Based on building information (BIM) of the office building and fire safety standards (such as GB500162014).

[0274] Specific operations:

[0275] Identify key points: safety exits, stairwells, and corridor intersections.

[0276] Generate standard evacuation routes: for example:

[0277] Path 1: Third floor office → Corridor L1 → Staircase S1 → First floor exit (90 meters in length).

[0278] Path 2: Third floor office → corridor L2 → staircase S2 → first floor exit (100 meters long).

[0279] Verify compliance with regulations: Ensure that the route meets the requirements for maximum evacuation distance and passage width.

[0280] Output standardized evacuation route data, including node sequences and route attributes.

[0281] (2) Hazard value acquisition module

[0282] The three floors of the office building were divided into a 1m×1m grid, generating multiple grid points.

[0283] Specific operations:

[0284] Normalize the environmental data for each grid point:

[0285] Temperature factor: min-max normalization is used (e.g., 50°C is normalized to 0.6).

[0286] Smoke concentration factor: min-max normalization is used (e.g., 200ppm is normalized to 0.7).

[0287] Personnel density factor: min-max normalization is used (e.g., 0.8 people / square meter is normalized to 0.8).

[0288] Calculate the overall hazard value of grid point G1 based on the weights (temperature weight 0.4, smoke concentration weight 0.4, and personnel density weight 0.2):

[0289] .

[0290] Output the overall hazard value for each grid point (e.g., G1: 0.66).

[0291] (3) Candidate path acquisition module

[0292] Dangerous grid points are removed, and candidate paths are generated using a reinforcement learning model.

[0293] Specific operations:

[0294] Dangerous grid points are removed: a preset danger threshold is 0.6, grid point G1 (danger value 0.66) is marked as impassable.

[0295] Constructing feasible space: remaining grid points (such as G2 danger value 0.5, G3 danger value 0.4) are used for path planning.

[0296] Input reinforcement learning decision model: encode grid point danger values as state space, model outputs action sequence (such as "move left", "move forward") based on policy network.

[0297] Generate candidate path sequence: for example:

[0298] C1: [start, G2, G4, G6,..., exit] (node sequence).

[0299] Verify path feasibility: check path connectivity.

[0300] Output candidate path sequence data, including node list and path length.

[0301] (4) Safety score calculation module

[0302] Evaluate the safety of the candidate path.

[0303] Specific operation:

[0304] Initial safety score calculation: match the node sequence of candidate path C1 with the standard path Path1 (based on edit distance algorithm), matching degree 80%, initial safety score 80.

[0305] Deduction adjustment: traverse the grid points of C1, for example:

[0306] G4 danger value 0.65, difference from threshold 0.6 is 0.05, deduct 5 points.

[0307] G6 danger value 0.55, no deduction.

[0308] G2 personnel density normalized value 0.9, preset density threshold 0.8, difference 0.1, deduct 10 points.

[0309] Adjusted safety score: 80-5-10=65.

[0310] Output path safety score data, including deduction details and safety level.

[0311] (5) Efficiency score calculation module

[0312] Evaluate the evacuation efficiency of the candidate path.

[0313] Specific operation:

[0314] Initial efficiency score calculation: C1 total length 100 meters, standard path length 90 meters, ratio 1.11, initial efficiency score 90.

[0315] Penalty adjustment: analyze path personnel density, for example:

[0316] G4 hazard value 0.65, difference from threshold value 0.6 is 0.05, deduct 5 points.

[0317] G2 personnel density normalized value 0.9, preset density threshold value 0.8, difference 0.1, deduct 10 points.

[0318] Adjusted efficiency score: 90-10-5=75.

[0319] Output path efficiency score data, including bottleneck analysis and efficiency level.

[0320] (6) Optimal path acquisition module

[0321] Select the optimal path from the candidate paths and guide evacuation.

[0322] Specific operation:

[0323] Multi-objective optimization: use a weighted comprehensive evaluation function, safety weight 0.6, efficiency weight 0.4, C1 comprehensive score:

[0324] .

[0325] Pareto optimal solution set screening: compare all candidate paths, C1 comprehensive score is the highest.

[0326] Real-time verification: combined with the latest environmental data, confirm that the path is not blocked by flames.

[0327] Send guidance instructions: convert optimal path data into control instructions and send to intelligent indication device (such as electronic indication board), display "evacuate left" arrow.

[0328] The technical scope of the present application is not limited to the content in the above description, and those skilled in the art can make various modifications and changes to the above embodiments without departing from the technical idea of the present application, and these modifications and changes should all belong to the protection scope of the present application.

Claims

1. An enhanced learning-driven fire emergency evacuation guidance system, characterized in that, include: The data acquisition module is used to collect environmental data and multiple standard evacuation routes at multiple predetermined locations within the building in real time. The environmental data includes temperature, smoke concentration, and personnel density. The hazard value acquisition module is used to divide the predetermined site into multiple grid points according to a preset size, normalize the environmental data in each grid point and perform weighted fusion processing to obtain a comprehensive hazard value. The candidate path acquisition module is used to remove grid points whose comprehensive hazard value is greater than a preset hazard threshold, and input the comprehensive hazard value of the remaining grid points after the removal process into the reinforcement learning decision model to generate a complete candidate path sequence. The safety score calculation module is used to calculate the matching degree between the node sequence of the candidate path sequence and the node sequence of the standard evacuation path to obtain an initial safety score, calculate the risk difference between the comprehensive risk value of each grid point in the path and the preset risk threshold, and calculate the density difference between the normalized personnel density and the preset density threshold. Based on the risk difference and density difference, the initial safety score is deducted and adjusted to obtain the path safety score. The efficiency score calculation module is used to calculate the ratio of the total length of the candidate path sequence to the preset standard path length to obtain an initial efficiency score, and to deduct points from the initial efficiency score according to the danger difference and density difference to obtain a path efficiency score. The optimal path acquisition module is used to select the optimal evacuation path from the candidate path sequence based on the path safety score and path efficiency score through a multi-objective optimization algorithm, and send the optimal evacuation path to the intelligent indicator device in the building for path guidance.

2. The reinforcement learning-driven fire emergency evacuation guidance system according to claim 1, characterized in that, Before normalizing and weighted fusion of environmental data at each grid point to obtain the comprehensive hazard value, a mechanism for verifying and completing the integrity of the environmental data is also included, specifically including: Acquire environmental data for each predetermined site and identify missing or abnormal data; Based on spatiotemporal correlation analysis, missing data is supplemented using interpolation algorithms with environmental data from adjacent grid points, and abnormal data is identified and corrected through data consistency checks.

3. The reinforcement learning-driven fire emergency evacuation guidance system according to claim 1, characterized in that, After normalizing and weighted fusion of the environmental data at each grid point to obtain the comprehensive hazard value, the method further includes a step of predicting and updating the comprehensive hazard value, specifically including: Continuously acquire historical time-series data of environmental data at grid points within the calibration time window; Based on the time series analysis algorithm, the historical time series data is analyzed and predicted to obtain a predicted value. When the predicted value is greater than the danger threshold, the grid point is marked as a danger diffusion area, and the comprehensive danger value of the grid point is replaced with the predicted value. The candidate path containing the grid point is removed from the candidate path sequence and the candidate path sequence is updated.

4. The reinforcement learning-driven fire emergency evacuation guidance system according to claim 1, characterized in that, The construction process of the reinforcement learning decision model specifically includes: Acquire historical environmental data of building grid points and make path selection decisions; Historical environmental data is abstracted into a discrete set of states to construct a state space; The path selection decision is abstracted into a discrete set of actions to construct an action space; A reward function is constructed based on safety rewards and efficiency rewards. The safety reward is determined based on the difference between the comprehensive hazard value of the path and the hazard threshold, while the efficiency reward is determined based on the ratio of the path length to the preset standard path length. A policy network is constructed by fitting the probability distribution mapping relationship from the state space to the action space through a deep neural network. A reinforcement learning decision model is constructed based on the reward function and policy network.

5. The reinforcement learning-driven fire emergency evacuation guidance system according to claim 1, characterized in that, After inputting the comprehensive hazard value of the remaining grid points after the elimination process into the reinforcement learning decision model to generate a complete candidate path sequence, the process also includes a step of dynamically optimizing the candidate path sequence, specifically including: Obtain the generated candidate path sequence and calculate the path coherence in the sequence; If the path coherence is less than the coherence threshold, the candidate path will be eliminated. Calculate the minimum personnel density that the path nodes in the remaining candidate paths can accommodate. If the minimum personnel density is less than the current personnel density, the candidate path is eliminated, and the remaining candidate paths are used as the optimized candidate path sequence.

6. The reinforcement learning-driven fire emergency evacuation guidance system according to claim 5, characterized in that, Diversity assurance is provided for the dynamically optimized candidate path sequence, specifically including: Calculate the path repetition rate of the path coverage area in the optimized candidate path sequence; When the path repetition is greater than a preset repetition threshold, a supplementary path is generated based on the candidate path sequence before optimization.

7. The reinforcement learning-driven fire emergency evacuation guidance system according to claim 1, characterized in that, The process of selecting the optimal evacuation path from a sequence of candidate paths using a multi-objective optimization algorithm specifically includes: Based on the preset path comprehensive evaluation function, the path safety score and path efficiency score are weighted and calculated to obtain the comprehensive score. A multi-objective optimization algorithm is used to solve for the Pareto optimal solution set. The path with the highest comprehensive score is selected as the candidate path. The feasibility of the candidate path is verified, and the path that passes the verification is taken as the optimal evacuation path.

8. The reinforcement learning-driven fire emergency evacuation guidance system according to claim 7, characterized in that, Continuous monitoring of the effectiveness of optimal evacuation routes includes: Real-time acquisition of environmental monitoring data for the optimal evacuation route, and prediction of the remaining safe time for the route based on the changing trends of the environmental monitoring data; When the remaining safe time is less than a preset time threshold, the optimal evacuation route is replanned.

9. The reinforcement learning-driven fire emergency evacuation guidance system according to claim 1, characterized in that, After sending the optimal evacuation route data to the intelligent guidance devices inside the building for route guidance, the system also includes a collaborative guidance mechanism for the intelligent guidance devices, specifically including: Acquire the layout, location, and status information of each intelligent indicator device within the building; Based on the optimal evacuation route and the layout of the equipment, a hierarchical guidance instruction sequence is generated, including the main guidance scheme and the backup guidance scheme. Monitor the status information of the intelligent indicator device and determine whether it is normal. When it is in a normal state, activate the main guidance scheme; when it is in an abnormal state, activate the backup guidance scheme.

10. The reinforcement learning-driven fire emergency evacuation guidance system according to claim 9, characterized in that, Real-time monitoring of the evacuation effectiveness following the coordinated guidance mechanism includes: Obtain the population density during the actual evacuation process; Calculate the density difference between the personnel density and the density threshold, and adjust the policy network of the reinforcement learning decision model based on the density difference; The candidate path sequence is replanned based on the adjusted model, the optimal evacuation path is updated based on the replanned candidate path sequence, and the path guidance of the intelligent indicator device is updated in real time.

Images