A method and system for dynamic monitoring and early warning of fire risks in urban underground passages

Abstract

The present application relates to a method and system for dynamic monitoring and early warning of fire risks in urban underground passages, which solves the problem that single indicator evaluation is difficult to reflect the real risk and fixed thresholds cannot adapt to complex working conditions. The method includes: establishing a four-level threshold library based on the risk value, and dynamically adjusting the threshold boundary in combination with the scene; continuously monitoring the status when the risk value is lower than the preset threshold; when the risk value exceeds the threshold in the current scene, triggering the corresponding early warning and pushing it to the command center, while coupling historical fire data with dynamic attribute matrix data to calculate the fire probability of each grid; combining computational fluid dynamics to simulate the smoke diffusion path and the structural fire resistance limit, quantifying the severity of the fire consequences, and using a fuzzy comprehensive evaluation method to fuse the fire probability of each grid and the severity of the fire consequences to generate a risk matrix, and marking high-level risk areas. The present application has the following technical effects: realizing dynamic monitoring and early warning of fire risks in urban underground passages, improving assessment accuracy and early warning reliability.

Description

Technical Field

[0001] The present invention relates to the field of urban underground passage safety monitoring and early warning, and in particular to a method and system for dynamic monitoring and early warning of fire risks in urban underground passages. Background Art

[0002] With the development of urban transportation, urban underground passages have assumed an important transportation function. However, due to their strong closedness and dense traffic, they have a high fire risk and are prone to cause serious consequences. Traditional passive fire management can no longer meet the needs, and intelligent fire risk monitoring and early warning technology is urgently needed.

[0003] Currently, fire risk assessments in urban underground tunnels primarily rely on sensors collecting single data points like temperature and smoke, then combine fixed grid divisions with empirical formulas or simple machine learning models to determine risk. Some solutions utilize blockchain to ensure data credibility. These methods rely on a single metric, lack dynamic grid division adjustments, fail to consider the coupled relationships between multiple factors like ventilation and traffic flow, and fail to incorporate changes in time, weather, and operating conditions.

[0004] Existing technologies have obvious defects: the rough grid division leads to inaccurate risk positioning, the single indicator assessment is difficult to reflect the real risk, the fixed threshold cannot adapt to complex working conditions, and there is a lack of inter-regional risk transmission analysis and data anomaly processing mechanism, resulting in insufficient accuracy and reliability of monitoring and early warning, which is difficult to meet the actual needs of urban underground tunnel fire prevention and control. Summary of the Invention

[0005] In order to achieve dynamic monitoring and early warning of fire risks in urban underground passages and improve assessment accuracy and early warning reliability, the present application provides a method and system for dynamic monitoring and early warning of fire risks in urban underground passages.

[0006] In the first aspect, the present application provides a method for dynamic monitoring and early warning of fire risks in urban underground passages, which adopts the following technical solutions:

[0007] A method for dynamic monitoring and early warning of fire risks in urban underground passages, comprising:

[0008] Based on the tunnel layout and traffic density, the grid is adaptively divided and the range is determined. A dynamic attribute matrix containing fire risk assessment related parameters is assigned to each grid and stored in the blockchain.

[0009] Integrate dynamic attribute matrix data, multi-source heterogeneous data and historical fire data to build a fire risk assessment index system, and output real-time fire risk values ​​through a dynamic risk assessment model;

[0010] Establish a four-level threshold library based on risk values, and dynamically adjust the threshold boundaries based on the scenario;

[0011] Continuously monitor the status when the risk value is below the preset threshold;

[0012] When the risk value exceeds the threshold in the current scenario, the corresponding warning is triggered and pushed to the command center. At the same time, historical fire data and dynamic attribute matrix data are coupled to calculate the fire probability of each grid;

[0013] Combining computational fluid dynamics to simulate smoke diffusion paths and structural fire resistance limits, the severity of fire consequences is quantified. A fuzzy comprehensive evaluation method is used to fuse the fire probability of each grid and the severity of fire consequences to generate a risk matrix, and high-risk areas are marked.

[0014] By adopting the above technical solutions, this method accurately captures the real-time status of each area of ​​the tunnel through dynamic grid division and attribute matrix construction; integrates multi-source data and historical cases, and uses the spatiotemporal convolutional neural network model to achieve accurate fire risk assessment; dynamically adjusts the risk threshold, timely warns and calculates the probability of fire, combines smoke diffusion and fire resistance limit to quantify the consequences, generates a risk matrix to accurately mark high-risk areas, and effectively improves the fire prevention and control capabilities of urban underground passages.

[0015] Optional, adaptive meshing includes:

[0016] Based on the layout of the tunnel infrastructure, the Delaunay triangulation algorithm is used to generate the initial large grid boundary with the spatial coordinates of the monitoring equipment as the vertices, ensuring that each large grid contains at least three independent monitoring nodes;

[0017] Use risk perception equipment to collect multi-source data in the tunnel in real time, including traffic flow and environmental parameters;

[0018] A deep learning-based traffic flow classification algorithm processes and analyzes collected multi-source data in real time, extracts traffic flow features to calculate traffic density, and identifies scene categories, including congestion, unobstructed traffic, breakdown parking, illegal parking, and traffic accident parking.

[0019] The grid size is automatically adjusted using preset rules based on traffic density and scene recognition results.

[0020] By adopting the above technical solution, the adaptive grid division method generates an initial large grid based on the tunnel layout and monitoring equipment coordinates. The deep learning algorithm processes multi-source data, accurately extracts traffic flow characteristics and identifies scene categories, and then automatically adjusts the grid size according to traffic density and scene results. This can achieve accurate division and dynamic monitoring of risks in different areas within the tunnel, providing a more accurate spatial basis for fire risk assessment and improving assessment accuracy and real-time performance.

[0021] Optionally, based on traffic density and scene recognition results, the grid size is automatically adjusted using preset rules, including:

[0022] Analyze whether there are high-risk events in the tunnel or whether the fire risk indicators in key areas exceed the preset safety threshold;

[0023] If yes, the MDP model is used to divide the state of the multi-source data, determine the grid adjustment action, and use the reward function that includes accuracy, equipment utilization and resource consumption to iteratively try action combinations through the reinforcement learning algorithm to learn the optimal strategy;

[0024] Automatically adjust the grid size based on the learned optimal strategy, and dynamically update the MDP model parameters and reward function based on the tunnel's real-time operation data and historical experience;

[0025] If not, the existing settings remain unchanged. The existing settings are to automatically adjust the grid size using preset rules based on traffic density and scene recognition results.

[0026] By employing these technical solutions, this method accurately identifies high-risk events and dynamically adjusts grid size using an MDP model and reinforcement learning algorithm to optimize accuracy, equipment utilization, and resource consumption. It dynamically updates model parameters based on real-time data and historical experience, enabling intelligent grid size adjustment. This improves the accuracy of fire risk assessments and resource utilization efficiency, enhancing the safety of urban underground passages.

[0027] Optionally, the method further includes steps after outputting the real-time fire risk value, specifically as follows:

[0028] A preliminary risk assessment is conducted based on the preset thresholds of single risk indicators. If any indicator exceeds the threshold, a Level 1 warning is triggered immediately. The risk grid is initially marked and the exceeding parameter and initial risk value R0 are recorded.

[0029] Based on the dynamic attribute matrix and historical fire data, the fire risk assessment index system is integrated to calculate the consistency between real-time data and historical data according to the preset weights;

[0030] If the degree of fit exceeds the preset threshold, the data is confirmed to be reliable, the risk value R0 is retained, and cross-regional verification is initiated;

[0031] Construct a regional risk propagation model with preset parameters, including the adjacent region influence coefficient K1 and the non-adjacent region influence coefficient K2. Based on the abnormal conditions of similar indicators in adjacent and non-adjacent regions and the preset parameters, the current grid risk value is dynamically adjusted using preset adjustment rules;

[0032] Using the Bayesian algorithm combined with DS evidence theory, weights are assigned according to the historical accuracy of the data source, and the triple verification results are integrated to output the final risk value;

[0033] Otherwise, the data completion mechanism will be activated to generate a revised risk value.

[0034] By employing the aforementioned technical solutions, this method achieves accurate fire risk assessment and early warning through multi-level verification and fusion techniques. Initial single-indicator early warnings quickly flag risks, the consistency of the fire risk assessment indicator system verifies data reliability, a cross-regional risk propagation model dynamically adjusts risk values, and the integration of the Bayesian algorithm and DS evidence theory yields triple verification results. This improves the accuracy and reliability of risk assessments, effectively reduces false alarm rates, and enhances urban underground tunnel fire prevention and control capabilities.

[0035] Optionally, the data completion mechanism is performed by the following steps:

[0036] Identify the grid spatial coordinates, real-time traffic density, and similar indicator data for a preset number of adjacent grids corresponding to the abnormal indicator, and construct a context vector containing spatial location, timestamp, and upstream and downstream traffic characteristics;

[0037] A conditional adversarial network model trained with data from the same operating conditions within a preset historical period is used to generate complementary data using the context vector as input.

[0038] Use the preset spatiotemporal consistency verification rules to verify the temporal continuity and spatial consistency of the completed data;

[0039] The verified completed data will replace the original abnormal values, re-enter the dynamic risk assessment model, and calculate the completed risk value in combination with historical fire data to correct the initial risk assessment results.

[0040] By adopting the above technical solution, the data completion mechanism constructs a context vector, uses a conditional generative adversarial network model to generate supplementary data, and replaces outliers after verification of spatiotemporal consistency. This mechanism can effectively correct the initial risk assessment results, improve the accuracy and reliability of risk assessment, ensure the stable operation of the fire risk assessment system, and enhance the precision of urban underground passage fire prevention and control. Optionally, the preset parameters include the adjacent area influence coefficient K1 and the non-adjacent area influence coefficient K2. The optimization steps for these preset parameters are as follows:

[0041] Use the preset LSTM model to generate future ventilation trends, extract the current grid's real-time wind direction vector, combine the real-time wind direction vector with the grid's spatial vector, and dynamically adjust the influence coefficient of adjacent areas;

[0042] In a preset time sliding window, the traffic density is exponentially weighted integrated, the integration result is input into a preset fuzzy logic controller, and the integration result is mapped into a dynamic modulation coefficient;

[0043] Multiplying the corrected adjacent region influence coefficient by the dynamic modulation coefficient to form the final adjacent region influence coefficient;

[0044] The Euclidean distance between the non-adjacent grid and the current grid is calculated. If the Euclidean distance between the non-adjacent grid and the current grid is less than the preset value, the influence coefficient of the non-adjacent area remains unchanged. Otherwise, the influence coefficient of the non-adjacent area decreases according to the preset exponential decay rule.

[0045] By employing these technical solutions, we optimized the parameters of the regional risk propagation model, used the LSTM model to predict ventilation trends, dynamically adjusted the influence coefficients of adjacent regions, and combined them with the dynamic modulation coefficient of traffic density to accurately reflect the characteristics of risk propagation. Furthermore, by adjusting the influence coefficients of non-adjacent regions based on Euclidean distance, we made the model more consistent with actual risk propagation patterns, improving the accuracy and dynamic adaptability of fire risk assessment.

[0046] Optionally, dynamically adjusting the adjacent region influence coefficient includes:

[0047] Extract the real-time wind direction of the current grid and the spatial coordinates of the adjacent grids, and calculate the angle between the connecting line and the wind direction;

[0048] If the angle is within the preset angle threshold, the influence coefficient of the adjacent area is enhanced according to the preset correction coefficient;

[0049] If it is headwind or exceeds the threshold angle, the influence coefficient will be weakened according to the preset ratio;

[0050] The actual distance between grids is calculated, and a preset exponential decay factor is introduced to perform spatial constraints on the corrected influence coefficient.

[0051] By employing this technical solution, the influence coefficients of adjacent regions are dynamically adjusted, taking into account wind angle and inter-grid distance, enhancing or weakening the influence through correction coefficients, and introducing exponential decay factors for spatial constraints. This method accurately reflects the impact of wind direction and distance on the spread of fire risk, making risk assessments more realistic, improving the scientific nature and accuracy of assessments, and enhancing the targeted nature of fire prevention and control.

[0052] Optionally, the method further includes steps after marking high-risk areas and extremely high-risk areas based on the risk value threshold interval, specifically as follows:

[0053] Using dynamic grids as nodes, ventilation attenuation, traffic congestion, and equipment failure are spatially and temporally aligned with historical disaster data. The probability of a chain reaction between fire and secondary disasters is quantified using a Bayesian network, and used as edge weights to construct a directed graph. A time decay function is introduced to adjust the timeliness of edge weights.

[0054] Through the analytic hierarchy process, the risk tolerance threshold is optimized and calculated by integrating personnel capacity, evacuation channel capacity, traffic flow, and meteorological factors. The difference between the threshold and the real-time risk value constitutes a dynamic safety margin, and the future state is predicted using a Markov chain.

[0055] Based on the disaster propagation directed graph and combined with Monte Carlo simulation, the diffusion paths of secondary disasters at multiple time nodes in the future are probabilistically deduced to generate a dynamic set of impact areas.

[0056] It integrates fire risk, secondary disaster prediction and safety margin status, overlays real-time traffic and personnel distribution, and forms a three-dimensional dynamic risk heat map to plan the optimal evacuation route and push warnings and plans.

[0057] By adopting the above technical solutions, the method quantifies the chain reaction probability of fire and secondary disasters through Bayesian networks, constructs a directed graph and introduces a time decay function to adjust the timeliness; uses the hierarchical analysis method to optimize the risk tolerance threshold, calculates the dynamic safety margin and predicts the future state; when the safety margin is lower than the warning curve, activates the emergency response, uses the blockchain to share the evacuation channel status, and dispatches drone monitoring; combines Monte Carlo simulation to predict the diffusion path of secondary disasters and generate a dynamic impact area set; integrates multi-dimensional data to form a three-dimensional dynamic risk heat map, plans the optimal evacuation route and pushes warnings and plans, effectively improving the prevention and control and emergency response capabilities of urban underground passage fires and secondary disasters.

[0058] Secondly, the present application provides a dynamic monitoring and early warning system for urban underground passage fire risks, which adopts the following technical solutions:

[0059] A dynamic monitoring and early warning system for fire risks in urban underground passages comprises a memory, a processor and a program stored in the memory and executable on the processor. The program can implement the dynamic monitoring and early warning method for fire risks in urban underground passages as described in the first aspect when loaded and executed by the processor. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a flow chart of a method for dynamic monitoring and early warning of fire risks in urban underground passages according to an embodiment of the present application.

[0061] Figure 2 It is a flowchart of adaptive grid division according to another embodiment of the present application. DETAILED DESCRIPTION

[0062] The present application is further described in detail below with reference to the accompanying drawings.

[0063] Reference Figure 1 , a method for dynamic monitoring and early warning of fire risks in urban underground passages disclosed in this application, comprising:

[0064] Step S100: Based on the tunnel layout and traffic density, the grid is adaptively divided and the range is determined. A dynamic attribute matrix containing fire risk assessment related parameters is assigned to each grid and stored in the blockchain.

[0065] The tunnel layout mentioned in this application refers to the tunnel infrastructure layout. Tunnel infrastructure layout refers to the spatial distribution of fixed facilities such as ventilation, fire protection, and lighting within the tunnel. The traffic density mentioned in this application refers to real-time traffic density, which refers to the number of vehicles passing through a certain section of the tunnel per unit time. Adaptive meshing dynamically adjusts the meshing based on the tunnel layout and traffic density. Parameters relevant to fire risk assessment include spatial coordinates, real-time ventilation field, traffic characteristics, and equipment coverage.

[0066] The acquisition methods are as follows: 1. Tunnel layout data: obtained through construction drawings and 3D laser scanning. 2. Traffic density data: obtained through traffic cameras and ground sensor loops.

[0067] The necessary process is outlined as follows:

[0068] 1. Grid division: Adaptive grid division method is used to dynamically adjust the grid size according to the layout of tunnel infrastructure and real-time traffic density. In areas with high traffic density or complex facilities, the grid division is finer; in areas with low traffic density or simple facilities, the grid division is coarser. For detailed processing, please refer to Figure 2 Steps S110 to S140.

[0069] 2. Dynamic attribute matrix construction: Each grid is assigned a dynamic attribute matrix, including spatial coordinates, real-time ventilation field, traffic flow characteristics, and equipment coverage. This data is updated in real time to ensure the matrix reflects the current state.

[0070] 3. Data storage: The dynamic attribute matrix is ​​stored in the blockchain, and the tamper-proof nature of the blockchain is used to ensure data security and reliability.

[0071] Step S200 , integrating dynamic attribute matrix data, multi-source heterogeneous data and historical fire data, constructing a fire risk assessment index system, and outputting a real-time fire risk value through a dynamic risk assessment model.

[0072] Among them, the dynamic attribute matrix: a grid feature matrix containing spatial coordinates, real-time ventilation fields, traffic flow characteristics and equipment coverage. Multi-source heterogeneous data: data from different sources and in different formats, such as sensor data, historical records, real-time monitoring data, etc. The fire risk assessment index system: includes evaluation indicators in multiple dimensions such as ventilation attenuation coefficient, traffic congestion index, and equipment failure probability. The dynamic risk assessment model in this application is preset, which is constructed based on the spatiotemporal convolutional neural network. The spatiotemporal convolutional neural network (ST-CNN) is a deep learning model used to process spatiotemporal data that can capture feature changes in time and space. Multi-source heterogeneous data: integrates sensor data in the tunnel (such as temperature, smoke sensors), traffic data (traffic density, speed), historical fire case data, etc.

[0073] The necessary process is outlined as follows:

[0074] 1. Data Fusion: Combine dynamic attribute matrices with multi-source heterogeneous data to form a complete input dataset. Ensure data quality through data cleaning and preprocessing.

[0075] 2. Construction of a Fire Risk Assessment Index System: Based on the integrated data, a fire risk assessment index system was constructed, including ventilation attenuation coefficient, traffic congestion index, and equipment failure probability. Each indicator was calculated using a specific algorithm. For example, the ventilation attenuation coefficient was calculated using ventilation field data, and the traffic congestion index was calculated using traffic density and speed data.

[0076] 3. Dynamic Risk Assessment Model: A dynamic risk assessment model is constructed using a spatiotemporal convolutional neural network (ST-CNN). The model input is fire risk assessment indicator system data, and the output is a real-time fire risk value. ST-CNN can capture temporal and spatial feature changes to accurately assess fire risk.

[0077] 4. Risk value output: After the model runs, it outputs real-time fire risk value to provide a basis for subsequent early warning.

[0078] In step S300 , a four-level threshold library is established according to the risk value, and the threshold boundaries are dynamically adjusted in combination with the scenario.

[0079] Among them, the four-level threshold library: a database that stores the range of fire risk values ​​corresponding to different risk levels (low, medium, high, and extremely high). The scenarios mentioned in this application refer to the operating environment characteristics of underground passages constructed based on three types of scenario labels: time, weather, and working conditions. Among them, the time label covers time period information such as day and night, season, and weekdays; the weather label includes environmental parameters such as temperature, humidity, and wind speed; and the working condition label involves channel operating conditions such as traffic density and equipment operating status. Through the combination of these labels, the boundaries of the four-level threshold library can be dynamically adjusted to make the risk warning more in line with the actual scenario.

[0080] Scenario label data: Time, weather and working condition information are obtained in real time through the tunnel monitoring system, and scenario labels are generated by combining historical data.

[0081] The necessary process is roughly described as follows: 1. Initial threshold setting: Based on historical fire case data and expert experience, four risk thresholds of low, medium, high, and extremely high are set and stored in the risk threshold library. 2. Scenario label generation: Real-time acquisition of time, weather, and operating condition information, combined with historical data to generate time-weather-operating condition scenario labels. 3. Dynamic threshold adjustment: Based on the scenario label, the risk threshold is dynamically adjusted using preset dynamic adjustment rules (such as appropriately lowering the threshold on rainy days or during peak hours) to ensure that the threshold is adapted to the current scenario. 4. Threshold library update: The adjusted threshold is updated to the risk threshold library to provide an accurate reference for subsequent risk assessments.

[0082] Step S400: When the risk value is lower than a preset threshold, the status is continuously monitored.

[0083] Among them, the preset threshold is a critical value set according to the risk threshold library, which is used to determine whether the fire risk requires an early warning.

[0084] "Continuous monitoring when the risk value falls below the preset threshold" is the system's fundamental step in real-time, dynamic tracking of fire risks in urban underground tunnels. Specifically, when the real-time risk assessment value falls below the threshold for triggering an early warning, the system maintains normal monitoring mode, continuously collecting dynamic attribute matrix data (such as spatial coordinates, ventilation patterns, and traffic flow characteristics) for each grid within the tunnel. Simultaneously, it updates multi-source heterogeneous information and historical fire case data, providing a continuous data flow to support subsequent risk assessments.

[0085] Step S500: When the risk value exceeds the threshold value in the current scenario, the corresponding warning is triggered and pushed to the command center. At the same time, the historical fire data and dynamic attribute matrix data are coupled to calculate the fire probability of each grid.

[0086] The essential process is outlined as follows: 1. Risk Value Comparison and Alert Triggering: The real-time fire risk value output in step S200 is compared with the current scenario threshold set in step S300. If the risk value exceeds the threshold, the system immediately triggers an alert response of the corresponding level. 2. Alert Information Push: Alert information is pushed to the command center in real time, including the fire risk level, affected areas, and recommended measures.

[0087] The fire probability calculation method uses a Bayesian probability model, combining historical fire data stored on the blockchain with real-time data from the dynamic attribute matrix to dynamically update the fire probability of each grid. The specific steps are as follows: 1. Obtain historical data: Obtain from the blockchain the fire probability of grid A12 over the past year, which is 0.05. 2. Obtain real-time data: Obtain real-time data from the dynamic attribute matrix: high traffic density (impact factor 1.5); normal ventilation system (impact factor 1.0); low equipment coverage (impact factor 0.8). 3. Calculate the updated probability: Use the Bayesian model to update the fire probability: P(fire|data) = P(fire) × impact factor; prior probability P(fire) = 0.05; impact factor = 1.5 × 1.0 × 0.8 = 1.2; updated fire probability = 0.05 × 1.2 = 0.06; 4. Adjustment factor: Considering the impact of high traffic density, the adjustment factor is 1.5: P(fire|data) = 0.06 × 1.5 = 0.09. 5. Update to the blockchain: Store the updated fire probability of 0.09 on the blockchain.

[0088] Step S600 combines computational fluid dynamics to simulate the smoke diffusion path and the structural fire resistance limit, quantifies the severity of the fire consequences, and uses a fuzzy comprehensive evaluation method to fuse the fire probability and fire consequence severity of each grid to generate a risk matrix, and marks high-risk areas.

[0089] The risk matrix is ​​a matrix that assesses the fire risk of each grid by combining the probability of fire and the severity of fire consequences. The severity of fire consequences is the degree of casualties, property damage, and environmental damage that may result from a fire. The fuzzy comprehensive evaluation method is a multi-factor comprehensive evaluation method based on fuzzy mathematics, used to deal with uncertainty and ambiguity.

[0090] The necessary process is outlined as follows:

[0091] 1. Quantify the severity of fire consequences: Computational fluid dynamics (CFD) was used to simulate the smoke diffusion path and, combined with the structural fire resistance, quantify the severity of fire consequences. For example, according to the simulation results, the fire consequence severity of grid A12 was 0.8 (with a maximum score of 1 indicating very severe).

[0092] 2. Construct a risk matrix: Using the fuzzy comprehensive evaluation method, combine the fire probability (obtained from step S500) and the fire consequence severity to generate a risk matrix. For example, if the fire probability is 0.09 and the consequence severity is 0.8, the comprehensive risk value is calculated as 0.09 × 0.8 = 0.072.

[0093] 3. Mark high-risk areas: Based on the risk matrix and the preset risk threshold (e.g., a high-risk threshold of 0.07), mark high-risk areas. For example, the risk value of 0.072 in grid A12 exceeds the high-risk threshold and is marked as a high-risk area.

[0094] 4. Update the risk matrix to the blockchain: Storing the risk matrix on the blockchain ensures the data is tamper-proof and traceable, providing a basis for subsequent decision-making.

[0095] Reference Figure 2 , adaptive meshing includes:

[0096] In step S110 , based on the layout of the tunnel infrastructure, the Delaunay triangulation algorithm is used to generate the initial large grid boundary with the spatial coordinates of the monitoring equipment as vertices, ensuring that each large grid contains at least three independent monitoring nodes.

[0097] The Delaunay triangulation algorithm is a geometric algorithm used to segment a set of points into a triangular mesh, ensuring that the circumcircle of each triangle contains no other points. The spatial coordinates of monitoring equipment (such as sensors and cameras) installed in the tunnel are used. Independent monitoring nodes are monitoring equipment points with independent data collection and transmission capabilities.

[0098] The necessary process is briefly described:

[0099] 1. Initial grid boundary generation: Use the Delaunay triangulation algorithm to generate the initial large grid boundary with the spatial coordinates of the monitoring equipment as the vertices. Ensure that each large grid contains at least three independent monitoring nodes to ensure data reliability and coverage.

[0100] 2. Algorithm Application: The Delaunay triangulation algorithm automatically creates a triangular mesh based on the spatial coordinates of the monitoring devices, with each triangle vertex corresponding to a monitoring device. The algorithm ensures that the circumcircle of each triangle does not contain other monitoring device points, resulting in an efficient meshing process.

[0101] 3. Grid optimization: If there are less than 3 monitoring nodes in a grid, the grid boundaries will be automatically adjusted, and the grids will be merged or re-divided to ensure that each grid contains at least 3 independent monitoring nodes.

[0102] Example: Suppose there are 10 monitoring devices installed in a tunnel, and their spatial coordinates are known. Using the Delaunay triangulation algorithm, these devices are divided into several triangular meshes. Each mesh contains at least three monitoring devices to ensure comprehensive and reliable data collection. For example, devices A, B, and C form one triangular mesh, devices D, E, and F form another, and so on.

[0103] Step S120: Using risk perception equipment to collect multi-source data in the tunnel in real time, wherein the multi-source data includes traffic flow and environmental parameters.

[0104] Risk perception equipment: devices that can monitor multiple risk-related data in real time within tunnels, such as traffic flow sensors and environmental parameter sensors. Multi-source data: data from different types of sensors, including traffic flow data (vehicle speed, volume) and environmental parameters (temperature, humidity, and air quality).

[0105] In step S130, a deep learning-based traffic flow classification algorithm is used to process and analyze the collected multi-source data in real time, extract traffic flow features to calculate traffic flow density, and identify scene categories, including congestion, unobstructed traffic, parking due to breakdown, illegal parking, and parking due to traffic accidents.

[0106] The deep learning traffic flow state classification algorithm is a deep learning-based algorithm used to analyze traffic flow data and identify traffic flow states. Traffic flow features are features extracted from traffic flow data, such as vehicle speed, volume, and density, used to describe traffic flow states. Scenario categories are traffic flow state classification results, including congestion, unimpeded traffic, breakdown parking, illegal parking, and traffic accident parking.

[0107] The necessary process is outlined as follows:

[0108] 1. Data preprocessing: The traffic flow data collected in step S120 is preprocessed, including data cleaning (removing outliers), data normalization (converting the data to a uniform range), and feature extraction (extracting features such as vehicle speed, traffic volume, and vehicle density).

[0109] 2. Data cleaning rules: For example, remove abnormal data with a speed of 0 and a duration that is too long.

[0110] 3. Feature extraction: Extract key features from traffic flow data, such as vehicle speed, volume, and density. These features serve as input to the deep learning model. Feature extraction methods include, for example, calculating the average speed and volume at each monitoring point, as well as vehicle density (the number of vehicles per unit length).

[0111] 4. Deep Learning Model Application: Use a pre-trained deep learning model (such as a CNN or RNN) to classify the extracted features and identify the traffic flow scenario category. Model Selection: For example, use a convolutional neural network (CNN) to process time series data and identify traffic flow status. Classification Results: The model outputs the traffic flow scenario category, such as congestion, unimpeded traffic, and abnormal parking.

[0112] 5. Result output: The identified traffic flow scene categories are output to provide a basis for subsequent grid size adjustment.

[0113] Step S140 , automatically adjusting the grid size using preset rules according to the traffic density and scene recognition results.

[0114] Grid size adjustment rules: Preset specific rules for dynamically adjusting grid size based on traffic density and scenario category. Traffic density: The number of vehicles passing through a certain cross-section of a tunnel per unit time, reflecting traffic congestion. Scenario category: The classification of traffic conditions, including congestion, unimpeded traffic, and abnormal parking.

[0115] The necessary process is outlined as follows:

[0116] 1. Obtain traffic density and scene category: Obtain the traffic density and scene category (e.g., congested, unobstructed, or abnormal parking) for each grid from step S130. For example, the traffic density of grid A12 is 80 vehicles / km, and the scene category is "congested."

[0117] 2. Apply grid sizing rules:

[0118] The grid size is automatically adjusted based on preset rules. The rules are as follows: Congestion: The grid size is reduced to more precisely monitor high-risk areas. Unobstructed traffic: The grid size remains unchanged or is moderately expanded to reduce computing resource consumption. Abnormal traffic: The grid size is reduced, and the monitoring frequency is increased to quickly respond to potential risks. For example: For congested grid A12, the grid size is adjusted from 100m x 100m to 50m x 50m.

[0119] 3. Update Grid Size: Update the adjusted grid size to the system to ensure that subsequent steps use the latest grid division. Example: The updated grid A12 size is 50m x 50m. The system automatically adjusts the relevant monitoring and evaluation parameters.

[0120] Based on traffic density and scene recognition results, the grid size is automatically adjusted using preset rules, including:

[0121] Step S141: Analyze whether there are any high-risk events in the tunnel or whether the fire risk index of the key area exceeds the preset safety threshold.

[0122] High-risk events are defined as those that could potentially cause fires or other safety incidents, such as vehicle collisions and unusual parking situations. Critical areas are defined as safety-critical areas within tunnels, such as those near ventilation openings and areas where firefighting equipment is concentrated. Fire risk indicators are quantitative metrics used to assess fire risk, such as traffic density, ambient temperature, and smoke concentration. Pre-set safety thresholds are defined as critical safety values ​​set based on historical data and expert experience. Exceeding these thresholds indicates a high risk.

[0123] The necessary process is outlined as follows:

[0124] 1. Risk Indicator Analysis: Analyze the fire risk indicators of various areas within the tunnel to determine whether there are high-risk events or whether the fire risk indicators in key areas exceed preset safety thresholds. For example, check whether the traffic density in grid A12 exceeds a threshold (e.g., 100 vehicles / km) or whether the ambient temperature exceeds a threshold (e.g., 30°C).

[0125] 2. High-risk determination: If any fire risk indicator exceeds a preset safety threshold, or a high-risk event (such as an abnormal parking situation) occurs, a high-risk state is determined. For example, the traffic density in grid A12 is 120 vehicles / km, exceeding the preset threshold of 100 vehicles / km, resulting in a high-risk state.

[0126] 3. Result output: Output the high-risk judgment result to provide a basis for subsequent steps. Example: The output result is "Grid A12 has a high-risk event and the grid size needs to be adjusted."

[0127] For example, assume that the traffic density in tunnel grid A12 is 120 vehicles / km, the ambient temperature is 32°C, and the preset safety thresholds are 100 vehicles / km and 30°C, respectively. Analysis reveals that both the traffic density and the ambient temperature exceed the preset safety thresholds, determining that grid A12 has a high-risk event. The system outputs a high-risk determination and proceeds to step S142 to determine the grid adjustment action.

[0128] Step S142: If yes, the MDP model is used to divide the state of the multi-source data, determine the grid adjustment action, and use the reward function including accuracy, equipment utilization and resource consumption to iteratively try the action combination through the reinforcement learning algorithm to learn the optimal strategy.

[0129] The Markov Decision Process (MDP) model is a mathematical model used to describe the decision-making process in a random environment. It consists of states, actions, and reward functions. State represents the current state of the system, such as traffic density and environmental parameters. Actions represent the decisions the system can take, such as adjusting the grid size or increasing monitoring frequency. Reward function measures the quality of each action and typically includes factors such as accuracy, device utilization, and resource consumption.

[0130] The necessary process is outlined as follows:

[0131] 1. State Classification: Using the MDP model, we integrate multiple data sources (such as traffic density and environmental parameters) to classify states. For example, we classify states into three levels based on traffic density: low, medium, and high. For example, a traffic density of less than 50 vehicles / km is considered low, 50-100 vehicles / km is considered medium, and greater than 100 vehicles / km is considered high.

[0132] 2. Determine the action: Based on the partitioned state, determine possible grid adjustment actions. For example, in a high state, a possible action is to reduce the grid size. Example: In a high traffic density state, the action is to adjust the grid size from 100m x 100m to 50m x 50m.

[0133] 3. Reward Function Design: Design a reward function that incorporates accuracy, device utilization, and resource consumption. For example, actions with high accuracy, high device utilization, and low resource consumption receive higher rewards. Example: The reward function can be expressed as R = w1 × accuracy + w2 × device utilization − w3 × resource consumption, where w1, w2, and w3 are weights.

[0134] 4. Application of reinforcement learning algorithms: Reinforcement learning algorithms (such as Q-learning or DQN) are used to iteratively try combinations of actions to learn the optimal policy. Through trial and error, the algorithm evaluates the performance of each action based on a reward function, gradually learning the optimal grid adjustment strategy. For example, using the Q-learning algorithm, the system tries different grid adjustment actions, records the reward value for each action, and ultimately learns that in conditions of high traffic density, reducing the grid size yields the highest reward.

[0135] For example, assume that grid A12 has a traffic density of 120 vehicles per kilometer and an ambient temperature of 32°C, placing it in a high state. According to the MDP model, a possible action is to adjust the grid size from 100 m x 100 m to 50 m x 50 m. The reward function is designed as R = 0.5 × accuracy + 0.3 × device utilization - 0.2 × resource consumption. Using the Q-learning algorithm, the system learns that in high traffic density states, reducing the grid size yields the highest reward.

[0136] In step S143 , the grid size is automatically adjusted according to the learned optimal strategy, and the MDP model parameters and reward function are dynamically updated based on the real-time operation data and historical experience of the tunnel.

[0137] The necessary process is outlined as follows:

[0138] 1. Apply the optimal strategy: Automatically adjust the grid size based on the optimal strategy learned in step S142. For example, if the strategy indicates a smaller grid size under high traffic density conditions, the system will adjust the grid accordingly. Example: The optimal strategy indicates that grid A12 should be adjusted from 100m x 100m to 50m x 50m when traffic density is high.

[0139] 2. Dynamically update model parameters: The MDP model parameters and reward function are dynamically updated based on the tunnel's real-time operational data and historical experience. For example, if real-time data indicates that a certain adjustment strategy is ineffective in a specific scenario, the model parameters will be adjusted accordingly to optimize future decisions. For example, if real-time data indicates that grid A12 still presents a high risk after adjustment, the system will adjust the MDP model parameters based on this data and optimize the reward function weights to improve the effectiveness of the strategy.

[0140] 3. Continuous Optimization: The system continuously monitors the effects of adjustments and further optimizes the grid adjustment strategy based on new data and feedback to ensure its adaptability and effectiveness. Example: The system continuously monitors the traffic density and environmental parameters of grid A12 and dynamically adjusts the grid size based on new data to ensure risk control remains within a safe range.

[0141] As an example, assume that grid A12 has a traffic density of 120 vehicles per kilometer and an ambient temperature of 32°C, indicating a high-risk state. Based on the optimal strategy learned in step S142, the system adjusts the size of grid A12 from 100 m x 100 m to 50 m x 50 m. After this adjustment, the system dynamically updates the MDP model parameters and reward function based on real-time operational data and historical experience, ensuring continuous optimization of the strategy.

[0142] In step S144 , if the answer is no, the existing setting is maintained unchanged. The existing setting is to automatically adjust the grid size using a preset rule according to the traffic density and the scene recognition result.

[0143] Existing Settings: The system automatically adjusts the grid size based on traffic density and scene recognition results. Maintain: The current grid settings remain unchanged unless a high-risk event is detected or the fire risk indicator in a critical area does not exceed the preset safety threshold.

[0144] A method for dynamic monitoring and early warning of fire risk in urban underground passages further includes steps after outputting the real-time fire risk value, specifically as follows:

[0145] Step S201: Preliminary judgment of the risk value is made based on the preset single risk indicator threshold. If any indicator exceeds the standard, a first-level warning is immediately triggered, the risk grid is preliminarily marked, and the exceeding parameter and the initial risk value R0 are recorded.

[0146] The single risk indicator threshold is a critical value set for a specific risk indicator (such as traffic density or equipment failure probability). Exceeding this value is considered a risk exceeding the limit. Level 1 Warning: A preliminary alarm triggered when a risk indicator exceeds the limit, indicating a potential fire risk. Initial Risk Value (R0): The risk value recorded when the Level 1 Warning is triggered is used for subsequent risk assessment and verification.

[0147] The necessary process is outlined as follows:

[0148] 1. Preliminary Assessment: After the dynamic risk assessment model outputs a real-time fire risk value, the system immediately makes a preliminary assessment based on a preset single risk indicator threshold. Examples include checking whether the traffic density exceeds a threshold (e.g., 100 vehicles / km) or whether the probability of equipment failure exceeds a threshold (e.g., 10%).

[0149] 2. Triggering a Level 1 Alert: If any indicator exceeds the standard, the system immediately triggers a Level 1 alert, preliminarily marks the risk grid, and records the exceeded parameter and initial risk value R0. For example, if the traffic density is 120 vehicles / km, exceeding the preset threshold of 100 vehicles / km, a Level 1 alert is triggered, the risk grid is marked, and the initial risk value R0 is recorded as 0.7.

[0150] 3. Record Exceeding Parameters and Initial Risk Value R0: The system records the specific parameters that exceed the standard (such as traffic density and equipment failure probability) and the corresponding initial risk value R0, providing basic data for subsequent fire risk assessment indicator system verification and cross-regional verification. For example, if the traffic density is 120 vehicles / km and the equipment failure probability is 12%, the initial risk value R0 is 0.7.

[0151] In step S202 , based on the dynamic attribute matrix and historical fire data, the fire risk assessment index system is integrated and the degree of consistency between the real-time data and the historical data is calculated according to preset weights.

[0152] Among them, the degree of consistency is the degree of similarity between real-time data and historical data, which is used to evaluate the reliability of data and the similarity of risks.

[0153] The necessary process is outlined as follows:

[0154] 1. Data Fusion: Based on a dynamic attribute matrix and historical fire data, a fire risk assessment indicator system is integrated, including ventilation attenuation, traffic congestion, and equipment failure. For example, in real-time data, the traffic density is 120 vehicles / km, the ventilation attenuation coefficient is 0.8, and the equipment failure probability is 12%. In historical data, a similar scenario has a traffic density of 110 vehicles / km, a ventilation attenuation coefficient of 0.75, and an equipment failure probability of 10%.

[0155] 2. Calculate the degree of consistency: Calculate the degree of consistency between real-time data and historical data using preset weights. The weights are set based on the importance of each indicator and the reliability of the historical data.

[0156] Example: The weights are set to 0.4 for traffic density, 0.3 for ventilation attenuation coefficient, and 0.3 for equipment failure probability. The calculation result is: Degree of Fit = 0.4 × 0.917 + 0.3 × 0.938 + 0.3 × 0.833 = 0.89.

[0157] 3. Result output: Output the goodness of fit value to provide data support for subsequent steps. Example: Output goodness of fit is 0.89, indicating that the real-time data is highly consistent with the historical data.

[0158] In step S203, if the degree of fit exceeds the preset threshold, the data is confirmed to be reliable, the risk value R0 is retained, and cross-region verification is initiated.

[0159] Data reliability confirmation: By comparing the consistency between real-time data and historical data, confirm whether the data is reliable and provide a basis for subsequent risk assessment.

[0160] Step S204: construct a regional risk propagation model with preset parameters, including the adjacent region influence coefficient K1 and the non-adjacent region influence coefficient K2. Based on the abnormal conditions of similar indicators in adjacent and non-adjacent regions and the preset parameters, the current grid risk value is dynamically adjusted using preset adjustment rules.

[0161] Regional risk propagation model: A model used to simulate the spread of fire risk between different areas of the tunnel. Adjacent area influence coefficient (K1): Indicates the degree to which adjacent areas affect the risk of the current area.

[0162] Non-adjacent area impact coefficient (K2): indicates the degree of impact of non-adjacent areas on the risk of the current area.

[0163] The necessary process is outlined as follows:

[0164] 1. Model Construction: Build a regional risk propagation model, presetting the adjacent region influence coefficient K1 and the non-adjacent region influence coefficient K2. For example, setting K1 to 0.5 means that the adjacent region's risk has a 50% impact on the current region; setting K2 to 0.2 means that the non-adjacent region's risk has a 20% impact on the current region.

[0165] 2. Parameter setting: Based on the tunnel layout and historical fire data, set other parameters in the model, such as the risk propagation speed and the distance between zones. Example: Set the risk propagation speed to 10 meters per minute and the average distance between zones to 50 meters.

[0166] 3. Model validation: Use historical fire data to verify the accuracy of the model and adjust model parameters to improve prediction accuracy. Example: Through verification of historical fire cases, adjust the values ​​of K1 and K2 to ensure that the model can accurately reflect the risk propagation situation.

[0167] The specific process of dynamically adjusting the current grid risk value using preset adjustment rules is as follows: if abnormalities in similar indicators are detected in a preset number of adjacent grids, the current grid risk value is adjusted to R0 plus the product of the adjacent area impact coefficient and the preset risk increment; when the indicators in non-adjacent areas are abnormal, the risk values ​​are superimposed according to the product of the preset impact coefficient K2 and the preset risk increment, among which abnormalities in similar indicators include traffic density exceeding a preset threshold, equipment failure probability exceeding a preset percentage, and ventilation attenuation coefficient exceeding a preset threshold.

[0168] Among them, similar indicator anomalies refer to situations where indicators such as traffic density, equipment failure probability, and ventilation attenuation coefficient exceed preset thresholds. Preset risk increment: The risk value increase set based on historical data and expert experience is used to adjust the risk value of the current grid.

[0169] The necessary process is outlined as follows:

[0170] Detect abnormalities in adjacent grids: Detect whether a preset number of adjacent grids have abnormalities of the same indicators.

[0171] Example: It is detected that the traffic density of the adjacent grid A13 exceeds the preset threshold and the probability of equipment failure exceeds the preset percentage.

[0172] Adjust the current grid risk value: If abnormalities of similar indicators appear in adjacent grids, the current grid risk value is adjusted to the initial risk value R0 plus the product of the adjacent area influence coefficient K1 and the preset risk increment.

[0173] Example: The initial risk value R0 is 0.7, the adjacent area influence coefficient K1 is 0.5, and the preset risk increment is 0.1. The adjusted risk value is 0.7+0.5×0.1=0.75.

[0174] Handling non-adjacent grid anomalies: If non-adjacent grids exhibit similar indicator anomalies, the risk value is superimposed by multiplying the non-adjacent area impact coefficient K2 by the preset risk increment. For example, if the ventilation attenuation coefficient of non-adjacent grid A14 exceeds the preset threshold, the non-adjacent area impact coefficient K2 is 0.2, and the preset risk increment is 0.1, the risk value is superimposed by 0.2 × 0.1 = 0.02, and the final risk value is 0.75 + 0.02 = 0.77.

[0175] In step S206, the Bayesian algorithm is combined with the DS evidence theory to assign weights according to the historical accuracy of the data source, and the triple verification results are integrated to output the final risk value.

[0176] The Bayesian algorithm is a statistical method based on Bayes' theorem used to calculate the probability of an event. DS Evidence Theory is a mathematical theory for handling uncertainty and ambiguity, used to integrate evidence from different sources. The historical accuracy of each data source is used to assign weights based on historical data. The triple verification results include single-indicator trigger warnings, verification of the consistency of the fire risk assessment indicator system, and cross-regional risk transmission.

[0177] The necessary process is outlined as follows:

[0178] 1. Weight Assignment: Based on the historical accuracy of the data source, a Bayesian algorithm combined with DS evidence theory is used to assign preset weights to the triple verification results of single indicator triggering warnings, fire risk assessment indicator system consistency verification, and cross-regional risk transmission. For example, the single indicator triggering warning weight is 0.3, the fire risk assessment indicator system consistency verification weight is 0.4, and the cross-regional risk transmission weight is 0.3.

[0179] 2. Evidence Fusion: Integrate the results of the three verifications and output the final risk value after evidence fusion. Example: The single indicator trigger warning value is 0.7, the fire risk assessment indicator system consistency verification value is 0.8, and the cross-regional risk transmission value is 0.75.

[0180] Final risk value = 0.3×0.7+0.4×0.8+0.3×0.75=0.765.

[0181] 3. Result output: Output the final risk value after evidence fusion to provide a basis for subsequent decision-making.

[0182] Example: Output final risk value is 0.765.

[0183] Step S207: Otherwise, the data completion mechanism is activated to generate a revised risk value.

[0184] The necessary process is outlined as follows:

[0185] 1. Activate the data completion mechanism: If, in step S206, the risk value after evidence fusion does not reach the preset reliability threshold, activate the data completion mechanism. Example: If the final risk value is 0.765, but the reliability threshold is 0.8, activate the data completion mechanism.

[0186] 2. Data Correction: Use a data completion algorithm and historical data to correct the current risk value. Example: Use linear interpolation to refer to risk values ​​of similar scenarios in historical data to correct the current risk value.

[0187] 3. Generate a revised risk value: The revised risk value is output as the final assessment result. Example: The revised risk value is 0.8, which meets the reliability requirements and is output as the final assessment result.

[0188] The data completion mechanism is performed through the following steps:

[0189] Step SA00: Identify the grid spatial coordinates, real-time traffic density, and similar indicator data of a preset number of adjacent grids corresponding to the abnormal indicator, and construct a context vector containing spatial location, timestamp, and upstream and downstream traffic characteristics.

[0190] The context vector is a vector containing information such as spatial location, timestamp, and upstream and downstream traffic characteristics, which is used to describe the context of abnormal indicators. Abnormal indicators are indicators that exceed preset thresholds or do not conform to normal patterns, such as traffic density and equipment failure probability.

[0191] The necessary process is outlined as follows:

[0192] 1. Identify abnormal indicators: Identify the grid coordinates, real-time traffic density, and similar indicator data for the three adjacent grids corresponding to the abnormal indicator. For example, if grid A12 has an abnormal traffic density of 150 vehicles / km, identify its spatial coordinates (x=100, y=200), the real-time traffic density of 150 vehicles / km, and the traffic densities of adjacent grids A11, A13, and A14, which are 120, 130, and 140 vehicles / km, respectively.

[0193] 2. Construct a context vector: Construct a context vector that includes the spatial location, timestamp, and upstream and downstream traffic features. Example: The context vector is [x=100,y=200,timestamp=2025−05−25T14:30:00,traffic density=150,upstream traffic density=120,downstream traffic density=130].

[0194] 3. Data integration: Integrate the identified anomaly indicators and their contextual information into a complete context vector to provide input for subsequent data completion. Example: The integrated context vector is [100, 200, 2025−05−25T14:30:00, 150, 120, 130].

[0195] In step SB00, a conditional generative adversarial network (CGN) model trained with data from the same operating conditions within a preset historical period is fed with a context vector to generate complementary data. A CGN is a variant of a generative adversarial network (GAN) that generates data that meets specific conditions based on conditional input. A context vector is a vector containing information such as spatial location, timestamp, and upstream and downstream traffic characteristics, used to describe the context of the abnormal indicator.

[0196] The necessary process is outlined as follows:

[0197] 1. Model input: Input the context vector constructed in step SA00 into the conditional generative adversarial network model.

[0198] Input context vector [100, 200, 2025−05−25T14:30:00, 150, 120, 130].

[0199] 2. Data Generation: Generate completion data using a conditional generative adversarial network model. The model uses adversarial training between the generator and the discriminator to generate completion data that meets contextual conditions. For example, the generator generates completion data based on the input context vector, such as a traffic density of 135 vehicles per kilometer.

[0200] 3. Model Validation: The discriminator verifies the authenticity of the generated data. The discriminator evaluates whether the generated data conforms to the distribution and context of historical data. For example, the discriminator verifies whether the generated traffic density of 135 vehicles / km conforms to the distribution and context of historical data.

[0201] Step SC00: Verify the temporal continuity and spatial consistency of the completed data using the preset spatiotemporal consistency verification rules.

[0202] Among them, spatiotemporal consistency verification rules are used to verify whether data is continuous and consistent in time and space. Temporal continuity is the continuity of data in the time series, ensuring that the data has no sudden changes in time. Spatial consistency is the rationality of data in space, ensuring that data in adjacent areas are similar.

[0203] The necessary process is outlined as follows:

[0204] 1. Temporal continuity verification: Check the continuity of the completed data in the time series. Ensure that there are no sudden changes between the completed data and the data at the previous and next time points. For example, verify that the completed traffic density of 135 vehicles / km is continuous with the data at the previous and next time points (such as 2:29 PM and 2:31 PM).

[0205] 2. Spatial consistency verification: Check the spatial rationality of the completed data. Ensure that the completed data is similar to the data in adjacent areas. For example, verify that the completed traffic density of 135 vehicles / km is consistent with the data in adjacent grids A11 (120 vehicles / km), A13 (130 vehicles / km), and A14 (140 vehicles / km).

[0206] 3. Verification result output: Output the verification result. If the completed data passes the verification, proceed to the next step; if not, regenerate the completed data. Example: If the verification passes, the output result is "Completed data passed verification."

[0207] For example, assume the completed data generated in step SB00 is a traffic density of 135 vehicles / km. Using the spatiotemporal consistency verification rules, this data is checked for temporal continuity (compared to data from previous and subsequent time points) and spatial rationality (compared to data from adjacent grids). The verification results indicate that the completed data is continuous with the data from previous and subsequent time points and consistent with data from adjacent grids, thus passing the verification.

[0208] In step SD00, the original abnormal value is replaced by the verified supplementary data, re-entered into the dynamic risk assessment model, and the supplemented risk value is calculated in combination with the historical fire data to correct the initial risk assessment result.

[0209] Verified Completed Data: Data that has been verified to be correct based on spatiotemporal consistency verification rules. Dynamic Risk Assessment Model: A model used to assess fire risk in real time, outputting a risk value based on a fire risk assessment indicator system. Completed Risk Value: A risk value obtained through reassessment using completed data, used to revise the initial risk assessment results.

[0210] The necessary process is outlined as follows:

[0211] 1. Data replacement: Replace the original abnormal value with the verified supplementary data. For example, the original abnormal traffic density of 150 vehicles / km is replaced by the supplementary data of 135 vehicles / km.

[0212] 2. Re-evaluate the risk value: Re-enter the replaced data into the dynamic risk assessment model and calculate the completed risk value based on historical fire data. Example: After re-entering the model and combining it with historical fire data, the completed risk value is 0.78.

[0213] 3. Revise the initial risk assessment results: Use the completed risk value to revise the initial risk assessment results.

[0214] Example: The initial risk value is 0.82, and the revised risk value is 0.78, which more accurately reflects the current fire risk.

[0215] Specifically, historical fire data plays a key auxiliary role in this process. The system first filters out historical data from the past 30 days that matches the current scenario based on tags such as time, meteorological conditions, and operating conditions. For example, it filters out historical fire risk data that also occurred during the morning rush hour congestion and rainy weather conditions. Due to the timeliness of the data, the historical data is weighted using an exponential decay function (e.g., a decay factor of λ = 0.01), reducing the weight of data from 10 days ago to the initial 36.8%. Subsequently, a Bayesian fusion algorithm is used to input the completed data into the model to calculate the real-time risk value, which is then weighted and fused with the processed historical fire probabilities for the same operating conditions.

[0216] Example: After the completed data is input into the model, the real-time risk value is 0.8, while the prior risk obtained from the statistical screening of historical data of the same working conditions is 0.6. The weights are dynamically assigned according to the timeliness of the data, with the real-time data weight w1=0.7 and the historical data weight w2=0.3. The completed risk value R_final=0.7×0.8+0.3×0.6=0.74 is calculated through the fusion formula R_final=w1×R_real+w2×R_hist.

[0217] The preset parameters include the adjacent area influence coefficient K1 and the non-adjacent area influence coefficient K2. The optimization steps of the relevant preset parameters are as follows:

[0218] In step Sa00, the preset LSTM model is used to generate future ventilation trends, extract the real-time wind direction vector of the current grid, combine the real-time wind direction vector with the grid space vector, and dynamically adjust the influence coefficient of the adjacent area.

[0219] Among them, the LSTM model (Long Short-Term Memory Network) is a special type of recurrent neural network (RNN) that can learn long-term dependencies and is suitable for predicting time series data. The real-time wind direction vector represents the real-time wind direction of the current grid, typically consisting of wind speed and wind direction. The grid space vector represents the position of the grid in space, typically consisting of coordinate points. The adjacent region influence coefficient (K1) indicates the degree of influence of adjacent regions on the risk of the current region.

[0220] The necessary process is outlined as follows:

[0221] 1. Generate Future Ventilation Trends: This function uses a pre-defined LSTM model to generate future ventilation trends. The LSTM model predicts future wind direction and speed based on historical ventilation data. For example, the LSTM model predicts that the wind direction will change from north to south, with a gradual increase in wind speed, over the next 10 minutes.

[0222] 2. Extract real-time wind direction vector: Extract the real-time wind direction vector of the current grid from the ventilation sensor.

[0223] Example: The real-time wind direction vector of the current grid A12 is [wind speed = 5m / s, wind direction = 180∘].

[0224] 3. Combine the real-time wind direction vector with the grid space vector: This combines the real-time wind direction vector with the grid space vector to dynamically adjust the adjacent area influence coefficient K1. For example, the space vector of the current grid A12 is [x=100,y=200], and the space vector of the adjacent grid A13 is [x=150,y=200]. The K1 value is adjusted based on wind direction and distance.

[0225] 4. Dynamically adjust the adjacent area influence coefficient: Dynamically adjust the adjacent area influence coefficient K1 based on wind direction and distance. If the wind direction is toward the adjacent grid, increase the K1 value; if the wind direction is away from the adjacent grid, decrease the K1 value. For example, if the wind direction is from north to south and grid A13 is on the south side, increase the K1 value; if grid A11 is on the north side, decrease the K1 value.

[0226] Step Sb00: perform exponential weighted integration on the traffic density within a preset time sliding window, input the integration result into a preset fuzzy logic controller, and map the integration result into a dynamic modulation coefficient.

[0227] Time sliding window: A fixed time range used to process time series data; the data within the window is used for calculations and analysis. Exponentially weighted integration: A weighted integration method that assigns higher weight to recent data to reflect its greater impact. Fuzzy logic controller: A fuzzy logic-based controller that maps input data to output control signals and is suitable for handling uncertainty and ambiguity. Dynamic modulation coefficient: A coefficient generated by a fuzzy logic controller based on the exponentially weighted integration results of traffic density, used to adjust the regional impact coefficient.

[0228] The necessary process is outlined as follows:

[0229] Traffic density exponential weighted integration: Performs an exponential weighted integration of traffic density within a preset time sliding window. Recent data is given a higher weight to reflect its greater impact.

[0230] For example, the sliding window is 10 minutes, and the traffic density data is [120, 130, 140, 150, 160, 170, 180, 190, 200, 210] vehicles per kilometer. The exponentially weighted integral results in 165 vehicles per kilometer.

[0231] 2. Input the fuzzy logic controller: Input the exponentially weighted integral result into the preset fuzzy logic controller. Example: Input the integral result of 165 vehicles / km into the fuzzy logic controller.

[0232] 3. Mapping to a dynamic modulation coefficient: The fuzzy logic controller maps the input integral result to a dynamic modulation coefficient. This coefficient is used to adjust the regional influence coefficient. Example: The fuzzy logic controller outputs a dynamic modulation coefficient of 0.8.

[0233] In step Sc00 , the corrected adjacent region influence coefficient is multiplied by the dynamic modulation coefficient to form a final adjacent region influence coefficient.

[0234] Revised adjacent region impact coefficient: The dynamically adjusted adjacent region impact coefficient is used to more accurately reflect the impact of adjacent regions on the current regional risk.

[0235] Dynamic modulation coefficient: A coefficient generated by a fuzzy logic controller based on the exponentially weighted integral result of traffic density, used to adjust the regional impact coefficient.

[0236] The necessary process is outlined as follows:

[0237] 1. Obtain the revised adjacent region influence coefficient: Obtain the revised adjacent region influence coefficient from step Sa00. Example: The revised adjacent region influence coefficient K1 is 0.6.

[0238] 2. Obtain the dynamic modulation coefficient: Obtain the dynamic modulation coefficient from step Sb00. Example: The dynamic modulation coefficient is 0.8.

[0239] 3. Calculate the final adjacent area influence coefficient: Multiply the corrected adjacent area influence coefficient by the dynamic modulation coefficient to obtain the final adjacent area influence coefficient. Example: The final adjacent area influence coefficient is 0.6 × 0.8 = 0.48.

[0240] Step Sd00, calculating the Euclidean distance between the non-adjacent grid and the current grid. If the Euclidean distance between the non-adjacent grid and the current grid is less than a preset value, the influence coefficient of the non-adjacent area remains unchanged. Otherwise, the influence coefficient of the non-adjacent area decreases according to a preset exponential decay rule.

[0241] Euclidean distance: The straight-line distance between two points, used to measure spatial proximity. Non-adjacent region influence coefficient (K2): Indicates the degree to which non-adjacent regions influence the risk of the current region. Exponential decay rule: A rule used to reduce the influence coefficient, where the influence coefficient decreases exponentially as distance increases.

[0242] The necessary process is outlined as follows:

[0243] 1. Calculate Euclidean distance: Calculate the Euclidean distance between non-adjacent grids and the current grid. Example: The spatial coordinates of the current grid A12 are [x=100,y=200], and the spatial coordinates of the non-adjacent grid A14 are [x=200,y=300].

[0244] The specific calculation of Euclidean distance is as follows:

[0245] , where d is the distance.

[0246] Determine whether the distance is less than the preset value: If the Euclidean distance between the non-adjacent grid and the current grid is less than the preset value, the influence coefficient K2 of the non-adjacent area remains unchanged. Example: The preset value is 150, and the current distance is 141.42, which is less than the preset value, so K2 remains unchanged.

[0247] 2. Apply the exponential decay rule: If the Euclidean distance is greater than the preset value, the influence coefficient K2 of the non-adjacent area is reduced according to the preset exponential decay rule. Example: The preset value is 150, and the current distance is 160, which is greater than the preset value. Assume that the exponential decay rule is , where α=0.01 and d is the distance. The calculation results are as follows:

[0248] .

[0249] Dynamically adjust the adjacent area influence coefficient including:

[0250] Step Sa10: extract the real-time wind direction of the current grid and the spatial coordinates of the adjacent grids, and calculate the angle between the direction of the line connecting the two and the wind direction.

[0251] Real-time wind direction: The wind direction of the current grid, typically measured in real time by a wind direction sensor. Adjacent grid spatial coordinates: The spatial coordinates of the adjacent grid, typically provided by tunnel layout data. Angle: The angle between the real-time wind direction and the line connecting the current grid and the adjacent grid.

[0252] Step Sa20: If the angle is within the preset angle threshold, the adjacent area influence coefficient is enhanced according to the preset correction coefficient.

[0253] The preset angle threshold is a range of angles set based on historical data and expert experience, used to determine the relationship between wind direction and the direction of the grid lines. The correction factor is a preset value used to increase or decrease the influence of adjacent areas.

[0254] The necessary process is outlined as follows:

[0255] 1. Determine whether the angle is within the preset threshold: If the angle calculated in step Sa10 is within the preset angle threshold, the adjacent area influence coefficient is enhanced according to the preset correction factor. Example: The preset angle threshold is [0°, 45°], and the angle calculated in step Sa10 is 30°, which is within the threshold.

[0256] 2. Enhance the influence coefficient of adjacent areas: Enhance the influence coefficient of adjacent areas according to the preset correction coefficient.

[0257] Example: The preset correction coefficient is 1.2, the initial adjacent area influence coefficient K1 is 0.5, and the enhanced K1 is 0.5×1.2=0.6.

[0258] 3. Result output: Output the enhanced adjacent region influence coefficient. Example: Output the enhanced adjacent region influence coefficient as 0.6.

[0259] Step Sa30: If it is a headwind or the angle exceeds the threshold, the influence coefficient is weakened according to a preset ratio.

[0260] Headwind refers to situations where the wind direction is opposite to the grid lines. Over-threshold angle refers to situations where the angle exceeds the preset angle threshold. Preset ratio refers to a preset value used to reduce the influence of adjacent areas.

[0261] The necessary process is outlined as follows:

[0262] 1. Determine whether the angle is headwind or exceeds the threshold: If the angle calculated in step Sa10 exceeds the preset angle threshold or is headwind, the influence coefficient of the adjacent area is weakened according to the preset ratio. For example, if the preset angle threshold is [0°, 45°] and the angle calculated in step Sa10 is 120°, which exceeds the threshold.

[0263] 2. Weaken the adjacent area influence coefficient: Weaken the adjacent area influence coefficient according to the preset ratio. Example: If the preset ratio is 0.8 and the initial adjacent area influence coefficient K1 is 0.5, the weakened K1 is 0.5 × 0.8 = 0.4.

[0264] 3. Result output: Output the weakened adjacent area influence coefficient. Example: Output the weakened adjacent area influence coefficient as 0.4.

[0265] In step Sa40 , the actual distance between grids is calculated, a preset exponential decay factor is introduced, and spatial constraints are applied to the corrected influence coefficients.

[0266] The actual distance between grids is the straight-line distance between the current grid and the adjacent grid. The exponential decay factor is the decay parameter used to adjust the influence coefficient based on the distance. The longer the distance, the smaller the influence.

[0267] The necessary process is outlined as follows:

[0268] 1. Calculate the actual distance between grids: Calculate the actual distance between the current grid and the adjacent grid based on their spatial coordinates.

[0269] 2. Introducing an exponential decay factor: Introducing a preset exponential decay factor, adjusting the corrected influence coefficient based on the actual distance between grids. Example: If the preset exponential decay factor is α = 0.01, the corrected adjacent area influence coefficient K1 is 0.6.

[0270] 3. Calculate the final impact coefficient: Calculate the final impact coefficient according to the exponential decay rule.

[0271] Example: The final impact coefficient calculation formula is: , where s represents the distance;

[0272] Substitute specific values:

[0273] .

[0274] 3. Result output: Output the final adjacent region influence coefficient. Example: Output the final adjacent region influence coefficient as 0.3639.

[0275] A method for dynamic monitoring and early warning of fire risk in urban underground passages further includes steps after marking high-risk areas and extremely high-risk areas based on risk value threshold intervals, specifically as follows:

[0276] In step S701, the dynamic grid is used as a node to align ventilation attenuation, traffic congestion, equipment failure, and historical disaster data in time and space. The probability of a chain reaction between fire and secondary disasters is quantified using a Bayesian network. This is used as the edge weight to construct a directed graph, and a time decay function is introduced to adjust the timeliness of the edge weight.

[0277] The Bayesian network is a probabilistic graphical model based on Bayes' theorem, used to represent conditional dependencies between variables. Edge weights represent the weight of connections between nodes in a directed graph and are used to quantify the probability of a chain reaction between fire and secondary disasters. Time decay is a function used to adjust the timeliness of edge weights, gradually reducing their influence over time.

[0278] The necessary process is outlined as follows:

[0279] 1. Spatiotemporal alignment: Using dynamic grids as nodes, align ventilation attenuation, traffic congestion, and equipment failure with historical disaster data. For example, align the ventilation attenuation coefficient of 0.8, traffic congestion index of 0.7, and equipment failure probability of 0.1 for the current grid A12 with historical disaster data.

[0280] 2. Quantify the probability of a chain reaction: Use a Bayesian network to quantify the probability of a chain reaction between a fire and secondary disasters, using this probability as the edge weight to construct a directed graph. Example: A Bayesian network calculates the probability of a fire causing a secondary disaster to be 0.6, using this as the edge weight to construct a directed graph.

[0281] 3. Introducing time decay function: Introducing time decay function to adjust the timeliness of edge weights and ensure that edge weights gradually weaken over time.

[0282] Example: The time decay function is , where β = 0.01 and time t = 10 minutes. The adjusted edge weight is 0.5429.

[0283] Output: Output the adjusted directed graph to provide a basis for subsequent steps. Example: Output the adjusted directed graph containing nodes A12 and A13, with an edge weight of 0.5429.

[0284] In step S702, the risk tolerance threshold is optimized and calculated by using the analytic hierarchy process, taking into account personnel capacity, evacuation channel capacity, vehicle flow, and meteorological factors. The difference between the threshold and the real-time risk value constitutes a dynamic safety margin, and the future state is predicted using a Markov chain.

[0285] Among them, the Analytic Hierarchy Process (AHP) is a method for multi-criteria decision analysis that determines the weight of each factor by constructing a hierarchical model. The risk tolerance threshold is the maximum risk value that the system can withstand, exceeding which requires action. The dynamic safety margin is the difference between the real-time risk value and the risk tolerance threshold, indicating the current safety level of the system. The Markov chain is a random process model used to predict future states based on the probability distribution of the current state.

[0286] The necessary process is outlined as follows:

[0287] 1. Comprehensive Factor Analysis: Using the Analytic Hierarchy Process (AHP), we comprehensively consider personnel capacity, evacuation channel capacity, traffic flow, and weather factors to optimize the risk tolerance threshold. For example, if the personnel capacity is 100, the evacuation channel capacity is 50 people per minute, the traffic density is 150 vehicles per kilometer, and the weather conditions are sunny, the AHP calculation results in a risk tolerance threshold of 0.8.

[0288] 2. Calculate the dynamic safety margin: The difference between the real-time risk value and the risk tolerance threshold constitutes the dynamic safety margin. Example: The real-time risk value is 0.6, the risk tolerance threshold is 0.8, and the dynamic safety margin is .

[0289] 3. Predicting future states: Use a Markov chain to predict future states based on the probability distribution of the current state. Example: If the current state is low risk, the Markov chain predicts that the probability of the state being low risk in the next 10 minutes is 0.9, and the probability of medium risk is 0.1.

[0290] 4. Result output: Output the dynamic safety margin and future state prediction results. Example: Output dynamic safety margin is 0.2, and the state prediction in the next 10 minutes is low risk (probability 0.9).

[0291] In step S703, when the safety margin is lower than the dynamic warning margin curve, an immediate response is initiated, emergency lighting and evacuation guidance are activated, the evacuation channel status is shared with the help of blockchain, and drones are dispatched to build a monitoring network.

[0292] Among them, the dynamic warning margin curve: a warning threshold curve set based on historical data and expert experience, used to dynamically determine whether a warning should be triggered. Emergency lighting and evacuation guidance: a lighting and sign system used to guide personnel evacuation in the event of a fire or other emergency. Blockchain-based shared evacuation channel status: using blockchain technology to share the status of evacuation channels in real time, ensuring the accuracy and immutability of information. Drone monitoring network: a monitoring network built using drones for real-time monitoring of fire scenes and evacuation operations.

[0293] The necessary process is outlined as follows:

[0294] 1. Determine whether the safety margin is lower than the warning margin curve: Compare the dynamic safety margin with the dynamic warning margin curve to determine whether a response needs to be initiated. Example: The dynamic safety margin is 0.2, and the dynamic warning margin curve is 0.3. The safety margin is lower than the warning margin curve.

[0295] 2. Initiate emergency response: When the safety margin falls below the dynamic warning margin curve, an immediate response is initiated, activating emergency lighting and evacuation guidance. Example: Activate emergency lighting and evacuation indicators in a tunnel to guide evacuation.

[0296] 3. Shared evacuation channel status: Using blockchain technology to share evacuation channel status ensures accuracy and immutability. Example: Using the blockchain platform to update the evacuation channel status in real time ensures that rescuers and evacuees have access to the latest information.

[0297] 4. Dispatch drones to build a monitoring network: Dispatch drones to build a monitoring network to monitor the fire scene and evacuation in real time. Example: Dispatch drones to fly to the fire area and transmit real-time video and data from the fire scene to assist in rescue decision-making.

[0298] 5. Result output: Output the emergency response activation status and related information. Example: Output that the emergency response has been initiated, emergency lighting and evacuation guidance have been activated, and the drone has been dispatched.

[0299] Step S704: Based on the disaster propagation directed graph and combined with Monte Carlo simulation, a probability deduction of the secondary disaster diffusion path at multiple time nodes in the future is performed to generate a dynamic impact area set.

[0300] Disaster propagation directed graph: A graph constructed based on the probability of chain reactions between fires and secondary disasters, used to represent the path of disaster propagation. Monte Carlo simulation: A statistical method that uses repeated random sampling to obtain numerical solutions, used to predict the probability distribution of future events. Dynamic impact area set: A set of areas that may be affected in the future, derived from the prediction of the disaster propagation path.

[0301] The necessary process is outlined as follows:

[0302] Obtaining a Disaster Propagation Directed Graph: Obtain a disaster propagation directed graph from step S701, which contains the probability of a chain reaction between fire and secondary disasters. Example: The directed graph contains nodes A12 and A13, with an edge weight of 0.5429, indicating the probability of fire propagating from A12 to A13.

[0303] Monte Carlo Simulation: Use Monte Carlo simulation to probabilistically predict the spread of secondary disasters at multiple time points in the future. Example: Run 1,000 simulations to predict the spread of disasters in the next 10, 20, and 30 minutes.

[0304] Generate a dynamic impact region set: Based on the Monte Carlo simulation results, a dynamic impact region set is generated to represent the areas that may be affected in the future. For example, it is predicted that within the next 10 minutes, grids A12 and A13 may be affected; within 20 minutes, grid A14 may also be affected.

[0305] Output: Output a set of dynamically affected areas to provide a basis for subsequent steps. Example: Output a set of dynamically affected areas, including the affected areas [A12, A13] within the next 10 minutes and the affected areas [A12, A13, A14] within the next 20 minutes.

[0306] Assume that the disaster propagation directed graph obtained in step S701 contains nodes A12 and A13, with an edge weight of 0.5429. Use 1000 Monte Carlo simulations to predict the disaster spread path for the next 10, 20, and 30 minutes. Based on the simulation results, a set of dynamic impact areas is generated: within the next 10 minutes, grids A12 and A13 may be affected; within 20 minutes, grid A14 may also be affected. Finally, the dynamic impact area set is output to provide a basis for subsequent evacuation and rescue efforts.

[0307] Step S705 integrates fire risk, secondary disaster prediction, and safety margin status, overlays real-time traffic and personnel distribution, forms a three-dimensional dynamic risk heat map, plans the optimal evacuation route, and pushes warnings and plans.

[0308] 3D Dynamic Risk Heat Map: A visualization tool that displays fire risk, secondary disaster predictions, and safety margin status in a 3D map. Optimal Evacuation Route: The safest evacuation route is calculated based on the risk heat map and real-time data.

[0309] The necessary process is outlined as follows:

[0310] 1. Data Fusion: Fusion of fire risk, secondary disaster predictions, and safety margin status, overlaid with real-time traffic and personnel distribution data. Example: Combining fire risk values, secondary disaster predictions, and safety margin status with real-time traffic flow and personnel distribution data.

[0311] 2. Generate a 3D dynamic risk heat map: This generates a 3D dynamic risk heat map to visually display risk distribution. Example: The generated heat map shows high-risk areas in red, medium-risk areas in yellow, and low-risk areas in green.

[0312] 3. Plan the optimal evacuation route: Plan the optimal evacuation route based on the three-dimensional dynamic risk heat map.

[0313] Example: Plan the optimal evacuation route from grid A12 to evacuation channel A, avoiding high-risk areas.

[0314] 4. Push warnings and plans: Push warning information and evacuation plans to relevant personnel and systems. Example: Push warning information and evacuation routes to the command center and personnel in the tunnel.

[0315] The steps for determining the optimal evacuation route can be as follows:

[0316] Step S7051: Construct a three-dimensional spatiotemporal model containing tunnel grids and evacuation exits. Dynamically load node edge weights with real-time traffic capacity (traffic flow, equipment status), safety risk weights (secondary disaster probability, safety margin), and time cost factors (congestion index function) to form a dynamically adjustable network infrastructure.

[0317] A three-dimensional spatiotemporal network model is constructed with tunnel dynamic grid cells (the smallest evaluation unit of the tunnel's physical space discretization, dynamically adjusted using the Delaunay triangulation algorithm combined with real-time traffic density) and key evacuation exits (including emergency passages, communication passages, and shaft exits, whose locations are preset based on tunnel design drawings) as nodes. The edges connecting the nodes represent traversable / travelable paths, and the edge weights are dynamically loaded according to the following rules:

[0318] Real-time traffic capacity: Millimeter-wave radar and video recognition equipment are used to collect traffic density in real time. IoT sensors obtain equipment status (fault signals from rolling doors, fans, etc.). Traffic capacity is calculated as follows: traffic rate = basic rate × (1-traffic density / congestion threshold) × equipment availability rate (the congestion threshold is 80% of the tunnel's designed traffic capacity. The equipment availability rate is determined in real time by the sensor status, with a normal value of 1 and a fault value of 0).

[0319] Safety risk weight: Directly call the secondary disaster occurrence probability output by the disaster propagation directed graph in step 8 (such as the path weight of explosion risk ≥ 70% × 2), and the safety margin assessment result (the grid adjacent path weight of safety margin < 0.3 times the risk tolerance threshold × 1.5, the risk tolerance threshold is calculated based on parameters such as personnel capacity and evacuation channel width).

[0320] Time cost factor: Based on 365 days of historical traffic flow data, a time period-congestion index mapping table is fitted and generated (for example, a congestion index of 1.3 from 7:00 to 9:00 on weekdays corresponds to a 30% increase in travel time; an index of 0.8 during non-peak hours corresponds to a 20% reduction in travel time). The data is loaded based on the current timestamp.

[0321] In step S7052, the improved NSGA-II algorithm is used to generate a Pareto optimal route solution that takes into account both efficiency and safety, with the goals of "shortest evacuation time, lowest disaster risk, and balanced channel load" and combined with constraints such as personnel movement speed and secondary disaster time window.

[0322] The improved non-dominated sorting genetic algorithm (NSGA-II) is used to generate a Pareto optimal route solution set based on the three optimization objectives of "minimizing evacuation time, minimizing disaster risk, and balancing channel load" and combining constraints. The specific process is as follows:

[0323] Input data preparation: Obtain node edge weight data (including real-time traffic capacity, safety risk weight, and time cost factor) from the three-dimensional space-time network model constructed in step S7051; extract the age distribution of people in the grid through the dynamic attribute matrix (used to calculate the movement speed of people, such as the average speed of 1.0m / s for the elderly and 1.5m / s for young and middle-aged people); retrieve the time window data for the secondary disaster diffusion prediction in step 8 (such as areas with high probability of explosion within the next 5 minutes).

[0324] Algorithm Execution and Constraints: A population of 100 initial evacuation routes was randomly generated, and optimization was performed iteratively through selection, crossover, and mutation. Constraints were set as follows: the instantaneous passenger load on a single route did not exceed 80% of the design capacity; and the routes avoided high-risk grids with a safety margin less than 0.2 times the risk tolerance threshold.

[0325] Fitness function calculation: The fitness of each route is calculated by comprehensively considering the weights of various objectives (evacuation time 40%, disaster risk 30%, and channel load 30%): evacuation time is accumulated according to the edge weight time cost factor; disaster risk takes the highest risk weight in the path; channel load is evaluated by the ratio of the traffic capacity of each edge of the path to the actual number of people carried, and finally non-dominated solutions are screened out to form the Pareto optimal solution set.

[0326] Step S7053 synchronizes real-time data from drone monitoring, vehicle-mounted equipment, and fire protection systems through blockchain, uses the YOLOv8 algorithm to dynamically identify channel congestion, temporary blocking points, and equipment status, updates network edge weights in real time, avoids high-risk paths, and implements dynamic route iteration.

[0327] A data synchronization mechanism is built based on blockchain technology to collect multi-source dynamic data in real time and drive iterative updates of evacuation routes. The specific process is as follows:

[0328] Real-time data collection and transmission: Drones equipped with high-definition cameras deployed in the tunnel use the YOLOv8 algorithm to analyze video streams in real time, identifying congestion and temporary blockage points in evacuation routes. Onboard OBUs report vehicle location information every five seconds, marking the coordinates of stalled vehicles. The firefighting IoT system provides real-time feedback on equipment operating status (e.g., sprinkler system failures, emergency lighting outages). All data is synchronized to a central processing system via a blockchain network, ensuring data immutability and consistency.

[0329] Dynamically update network edge weights: If the drone identifies a channel occupancy rate greater than 70%, the real-time capacity parameter of the corresponding path edge weight is reduced by 50%. When a vehicle's onboard equipment reports a vehicle stall point, the path it is on is automatically marked as impassable. If the fire system's feedback device fails, the corresponding regional path safety risk weight is increased by 50%. The network topology is recalculated based on the updated edge weight parameters.

[0330] Iterative route optimization: Based on the Pareto optimal solution set generated in step S7052, eliminate solutions that include impassable or high-risk paths. Use the Dijkstra algorithm to quickly select the new route with the lowest overall cost from the remaining solutions, enabling dynamic adjustment of the evacuation plan to ensure that the route always adapts to the real-time working conditions of the tunnel.

[0331] In step S7054, multimodal evacuation plans, including fast, fault-tolerant, and diversion plans, are generated for different groups of people. After the robustness of extreme scenarios is verified by the Unity digital twin scene, the optimal plan is encoded as machine-readable instructions and pushed through multiple terminals such as LED screens, apps, and in-vehicle systems, with real-time heat map navigation containing risk levels simultaneously marked.

[0332] A multimodal evacuation plan is generated based on differentiated needs and pushed to multiple terminals after verification by the digital twin. The specific process is as follows:

[0333] Multimodal solution generation: Output three types of solutions based on the characteristics of different groups of people:

[0334] Fast Track Solution: For people with unimpeded mobility, the shortest theoretical path is selected from the Pareto solution set in step S7052, while avoiding high-risk areas with a secondary disaster risk weight greater than 1.5.

[0335] Fault-tolerant channel plan: Designed for people with limited mobility, select an alternative route with a passage width ≥ 2 meters, no equipment failure, and a better time cost (no more than 15% increase in time cost);

[0336] Dynamic diversion plan: A user balance model is used to distribute evacuated people to various exits, ensuring that the instantaneous load of a single channel does not exceed 80% of the design capacity to avoid local congestion.

[0337] Digital twin verification and optimization: Using the Unity engine to build a 1:1 digital twin of the tunnel, the generated evacuation plan was imported into the simulation of extreme working conditions (such as sudden blockage of an exit or the triggering of a new secondary disaster). Crowd flow was simulated through a particle system to verify the feasibility of the plan in complex scenarios. Plans with path conflicts or sudden efficiency drops were iteratively optimized.

[0338] Multi-terminal command push: The optimal solution is encoded as navigation coordinates within the BIM model and pushed through multiple terminals, including the tunnel's LED guidance screen, mobile app, and in-vehicle central control system. A real-time risk heat map is simultaneously displayed, with prohibited areas marked in red, restricted areas in yellow, and safe areas in green, providing visual navigation guidance for evacuees.

[0339] Based on the same inventive concept, an embodiment of the present invention provides a dynamic monitoring and early warning system for fire risks in urban underground passages, including a memory and a processor. The memory stores data that can be executed on the processor to implement the following. Figures 1 to 2 Procedure for either method.

[0340] The embodiments of this specific implementation method are all preferred embodiments of the present application and are not intended to limit the scope of protection of the present application. Therefore, any equivalent changes made based on the structure, shape, and principle of the present application should be included in the scope of protection of the present application.

Claims

1. A method for dynamic monitoring and early warning of fire risk in urban underground passages, characterized in that: include: Based on the tunnel layout and traffic density, the grid is adaptively divided and the range is determined. A dynamic attribute matrix containing fire risk assessment related parameters is assigned to each grid and stored in the blockchain. Integrate dynamic attribute matrix data, multi-source heterogeneous data and historical fire data to build a fire risk assessment index system, and output real-time fire risk values ​​through a dynamic risk assessment model; Establish a four-level threshold library based on risk values, and dynamically adjust the threshold boundaries based on the scenario; Continuously monitor the status when the risk value is below the preset threshold; When the risk value exceeds the threshold in the current scenario, the corresponding warning is triggered and pushed to the command center. At the same time, historical fire data and dynamic attribute matrix data are coupled to calculate the fire probability of each grid; Combining computational fluid dynamics to simulate smoke diffusion paths and structural fire resistance limits, the severity of fire consequences is quantified. A fuzzy comprehensive evaluation method is used to integrate the fire probability and severity of each grid to generate a risk matrix, and high-risk areas are marked. Adaptive meshing includes: Based on the layout of the tunnel infrastructure, the Delaunay triangulation algorithm is used to generate the initial large grid boundary with the spatial coordinates of the monitoring equipment as the vertices, ensuring that each large grid contains at least three independent monitoring nodes; Use risk perception equipment to collect multi-source data in the tunnel in real time, including traffic flow and environmental parameters; A deep learning-based traffic flow classification algorithm processes and analyzes collected multi-source data in real time, extracts traffic flow features to calculate traffic density, and identifies scene categories, including congestion, unobstructed traffic, breakdown parking, illegal parking, and traffic accident parking. Automatically adjust the grid size using preset rules based on traffic density and scene recognition results; Based on traffic density and scene recognition results, the grid size is automatically adjusted using preset rules, including: Analyze whether there are high-risk events in the tunnel or whether the fire risk indicators in key areas exceed the preset safety threshold; If yes, the MDP model is used to divide the state of the multi-source data, determine the grid adjustment action, and use the reward function that includes accuracy, equipment utilization and resource consumption to iteratively try action combinations through the reinforcement learning algorithm to learn the optimal strategy; Automatically adjust the grid size based on the learned optimal strategy, and dynamically update the MDP model parameters and reward function based on the tunnel's real-time operation data and historical experience; If not, the existing settings remain unchanged. The existing settings are to automatically adjust the grid size using preset rules based on traffic density and scene recognition results.

2. A method for dynamic monitoring and early warning of fire risk in urban underground passages according to claim 1, characterized in that: The steps after outputting the real-time fire risk value are also included, as follows: A preliminary risk assessment is conducted based on the preset thresholds of single risk indicators. If any indicator exceeds the threshold, a Level 1 warning is triggered immediately. The risk grid is initially marked and the exceeding parameter and initial risk value R0 are recorded. Based on the dynamic attribute matrix and historical fire data, the fire risk assessment index system is integrated to calculate the consistency between real-time data and historical data according to the preset weights; If the degree of fit exceeds the preset threshold, the data is confirmed to be reliable, the risk value R0 is retained, and cross-regional verification is initiated; Construct a regional risk propagation model with preset parameters, including the adjacent region influence coefficient K1 and the non-adjacent region influence coefficient K2. Based on the abnormal conditions of similar indicators in adjacent and non-adjacent regions and the preset parameters, the current grid risk value is dynamically adjusted using preset adjustment rules; Using the Bayesian algorithm combined with DS evidence theory, weights are assigned according to the historical accuracy of the data source, and the triple verification results are integrated to output the final risk value; Otherwise, the data completion mechanism will be activated to generate a revised risk value.

3. A method for dynamic monitoring and early warning of urban underground passage fire risk according to claim 2, characterized in that: The data completion mechanism is performed through the following steps: Identify the grid spatial coordinates, real-time traffic density, and similar indicator data for a preset number of adjacent grids corresponding to the abnormal indicator, and construct a context vector containing spatial location, timestamp, and upstream and downstream traffic characteristics; A conditional adversarial network model trained with data from the same operating conditions within a preset historical period is used to generate complementary data using the context vector as input. Use the preset spatiotemporal consistency verification rules to verify the temporal continuity and spatial consistency of the completed data; The verified completed data will replace the original abnormal values, re-enter the dynamic risk assessment model, and calculate the completed risk value in combination with historical fire data to correct the initial risk assessment results.

4. A method for dynamic monitoring and early warning of urban underground passage fire risk according to claim 2, characterized in that: The preset parameters include the adjacent area influence coefficient K1 and the non-adjacent area influence coefficient K2. The optimization steps of the relevant preset parameters are as follows: Use the preset LSTM model to generate future ventilation trends, extract the current grid's real-time wind direction vector, combine the real-time wind direction vector with the grid's spatial vector, and dynamically adjust the influence coefficient of adjacent areas; In a preset time sliding window, the traffic density is exponentially weighted integrated, the integration result is input into a preset fuzzy logic controller, and the integration result is mapped into a dynamic modulation coefficient; Multiplying the corrected adjacent region influence coefficient by the dynamic modulation coefficient to form the final adjacent region influence coefficient; The Euclidean distance between the non-adjacent grid and the current grid is calculated. If the Euclidean distance between the non-adjacent grid and the current grid is less than the preset value, the influence coefficient of the non-adjacent area remains unchanged. Otherwise, the influence coefficient of the non-adjacent area decreases according to the preset exponential decay rule.

5. A method for dynamic monitoring and early warning of urban underground passage fire risk according to claim 4, characterized in that: Dynamically adjust the adjacent area influence coefficient including: Extract the real-time wind direction of the current grid and the spatial coordinates of the adjacent grids, and calculate the angle between the connecting line and the wind direction; If the angle is within the preset angle threshold, the influence coefficient of the adjacent area is enhanced according to the preset correction coefficient; If it is headwind or exceeds the threshold angle, the influence coefficient will be weakened according to the preset ratio; The actual distance between grids is calculated, and a preset exponential decay factor is introduced to perform spatial constraints on the corrected influence coefficient.

6. A method for dynamic monitoring and early warning of fire risk in urban underground passages according to any one of claims 1 to 5, characterized in that: The following steps are included after marking high-risk areas and very high-risk areas based on the risk value threshold interval: Using dynamic grids as nodes, ventilation attenuation, traffic congestion, and equipment failure are spatially and temporally aligned with historical disaster data. The probability of a chain reaction between fire and secondary disasters is quantified using a Bayesian network, and used as edge weights to construct a directed graph. A time decay function is introduced to adjust the timeliness of edge weights. Through the analytic hierarchy process, the risk tolerance threshold is optimized and calculated by integrating personnel capacity, evacuation channel capacity, traffic flow, and meteorological factors. The difference between the threshold and the real-time risk value constitutes a dynamic safety margin, and the future state is predicted using a Markov chain. Based on the disaster propagation directed graph and combined with Monte Carlo simulation, the diffusion paths of secondary disasters at multiple time nodes in the future are probabilistically deduced to generate a dynamic set of impact areas. It integrates fire risk, secondary disaster prediction and safety margin status, overlays real-time traffic and personnel distribution, and forms a three-dimensional dynamic risk heat map to plan the optimal evacuation route and push warnings and plans.

7. A dynamic monitoring and early warning system for urban underground passage fire risk, characterized in that: It includes a memory, a processor, and a program stored in the memory and executable on the processor. When the program is loaded and executed by the processor, it can implement a method for dynamic monitoring and early warning of fire risks in urban underground passages as described in any one of claims 1 to 6.

Images