Tunnel fire emergency linkage method and system based on digital twinning

Abstract

The application relates to the technical field of tunnel safety and emergency management, and discloses a tunnel fire-fighting emergency linkage method and system based on digital twinning, wherein the tunnel fire-fighting emergency linkage method based on digital twinning comprises the following steps: collecting multi-modal data of a fire environment through various sensors and constructing a global situation understanding model; adopting a multi-head attention network to process environmental state and crowd behavior characteristics; applying a context-enhanced attention mechanism to dynamically adjust the attention weight of the crowd characteristic dimension; constructing a psychological environment interaction model to predict the evolution of crowd behavior under different fire scenes; generating a context-adaptive multi-level evacuation guidance strategy tree; performing real-time effect evaluation, and realizing overall optimization of the system through a digital twinning model; the multi-head attention network can simultaneously process environmental state and crowd behavior characteristics, can understand the internal correlation between environmental factors and crowd behavior, and can form comprehensive and accurate cognition of a tunnel fire scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tunnel safety and emergency management, more particularly, it relates to a tunnel fire emergency linkage method and system based on digital twinning. BACKGROUND

[0002] With the acceleration of urbanization, tunnels are widely built and used as important transportation infrastructure. However, once a tunnel fire accident occurs, due to its enclosed space characteristics, it often causes serious casualties and property losses. At present, the tunnel fire emergency system still faces many technical challenges in practical application:

[0003] The existing technical solutions generally have the problem of single sensing ability, traditional systems usually only focus on the monitoring of fire physical parameters (such as temperature, smoke concentration, etc.), and cannot integrate fire situation and crowd psychological state information at the same time, which leads to incomplete cognition of the system to complex fire scenes, and the system cannot fully grasp the complex relationship between fire development and crowd reaction; most of the existing evacuation guidance strategies use preset schemes, which are out of touch with the actual situation, and these systems cannot provide differentiated guidance according to the dynamic changes of fire conditions and crowd psychological state in different areas, and cannot adapt to the uncertainty of fire development and the complexity of crowd behavior; the existing technology lacks understanding of the evolution law of crowd behavior in fire scenes, in the process of emergency evacuation, crowd behavior is affected by many factors, including environmental threat, individual psychological state and group effect, etc., and the prediction ability of the existing system to these complex behavior patterns is limited, which may lead to misjudgment at critical moments and affect the evacuation effect; the existing system generally lacks a collaborative sensing and analysis mechanism that can consider environmental factors and crowd psychological state at the same time, there is a complex interaction between fire environment and crowd psychological state, and traditional systems cannot effectively capture and utilize this interaction to optimize emergency decision-making.

[0004] The above technical problems lead to the fact that in complex tunnel fire evacuation scenarios, the existing system cannot provide accurate and effective emergency guidance, which increases the risk of casualties, and there is an urgent need for a new tunnel fire emergency linkage method that can consider both environmental situation and crowd psychological state. SUMMARY

[0005] The present application provides a tunnel fire emergency linkage method and system based on digital twinning, which solves the technical problems of being unable to collaboratively sense fire situation and crowd psychological state and being difficult to provide differentiated evacuation guidance in related technologies.

[0006] The present application provides a tunnel fire emergency linkage method based on digital twinning, comprising:

[0007] Collecting multi-modal data of fire environment by multiple sensors and constructing a global situation understanding model;

[0008] On the basis of the global situation understanding model, a multi-head attention network is used to process environmental state and crowd behavior characteristics to realize the collaborative perception of environment and crowd information.

[0009] Combined with the characteristics of fire scenes, a context-enhanced attention mechanism is applied to dynamically adjust the attention weight of crowd feature dimensions.

[0010] Based on the adjusted attention weight, a psychological environment interaction model is constructed to predict the evolution of crowd behavior under different fire scenarios.

[0011] According to the prediction results, a multi-level evacuation guidance strategy tree is generated to provide differentiated evacuation guidance.

[0012] During the implementation of the evacuation guidance strategy, real-time effect evaluation is carried out, and the overall system optimization is realized through the digital twin model.

[0013] Further, the global situation understanding model includes:

[0014] A multi-modal encoder unit is used to process different modal sensor data.

[0015] A cross-modal attention unit is used to calculate the correlation weight matrix between different modalities.

[0016] A feature fusion unit is used to weight and fuse each modal feature according to the attention weight.

[0017] A situation representation unit is used to convert the fused features into a physical quantity distribution representation of the tunnel space.

[0018] Further, the multi-head attention network includes:

[0019] An environmental feature input unit is used to receive and process environmental state features in the global situation understanding model.

[0020] A crowd feature input unit is used to receive and process crowd behavior characteristics.

[0021] A multi-head attention calculation unit includes Attention heads, each responsible for focusing on different environmental and crowd interaction patterns.

[0022] An attention output fusion unit is used to weight and fuse the outputs of Attention heads.

[0023] Further, the psychological environment interaction model includes:

[0024] An environmental perception unit is used to receive environmental state parameters.

[0025] a psychological state modeling unit configured to model the psychological state of the crowd;

[0026] a behavior evolution prediction unit configured to predict the change of the crowd behavior;

[0027] a danger state identification unit configured to identify the danger state.

[0028] Further, the multi-level evacuation guidance strategy tree comprises:

[0029] a root node unit representing the starting point of the overall evacuation decision;

[0030] an environmental state branch unit configured to divide the tunnel space into different regional nodes;

[0031] a psychological state branch unit configured to further branch according to the psychological state of the crowd;

[0032] a guidance strategy leaf node unit containing specific guidance strategies for specific combinations of environmental state and psychological state;

[0033] a priority scoring unit configured to assign priorities to each guidance strategy.

[0034] Further, the context-enhanced attention mechanism comprises:

[0035] acquiring fire scene characteristic data, including the location, scale, spread speed, and smoke diffusion state of the fire;

[0036] extracting and encoding the fire scene characteristics, converting them into a scene characteristic vector representation;

[0037] calculating attention weights for different crowd characteristic dimensions based on the scene characteristic vector;

[0038] applying the calculated attention weights to the multi-head attention network to dynamically adjust the attention degree to different crowd characteristic dimensions;

[0039] based on the adjusted attention distribution, scoring the importance of the crowd behavior characteristics, highlighting the most critical characteristics for evacuation guidance in the current scene.

[0040] Further, the real-time effect evaluation comprises:

[0041] defining guidance effect evaluation indicators, including evacuation speed, crowd flow smoothness, and panic emotion control degree;

[0042] collecting real-time data through a sensor network to monitor the execution effect of the guidance strategy;

[0043] using a bias function to calculate the deviation between the measured effect and the expected effect;

[0044] Generate a parameter adjustment amount according to the deviation analysis result.

[0045] Further, the digital twin model comprises:

[0046] A data synchronization unit is configured to receive and process real-time acquired situation data.

[0047] A virtual mapping unit is configured to map physical entity information to a virtual space.

[0048] A three-dimensional visualization unit is configured to display fire situation and crowd distribution.

[0049] A historical data storage unit is configured to save historical data of system operation.

[0050] A parameter optimization unit is configured to optimize parameters of each module of the system.

[0051] Further, the generated context-adaptive multi-level evacuation guidance strategy tree uses a multi-objective optimization algorithm, and the multi-objective optimization algorithm comprises:

[0052] Three optimization objectives, i.e., shortest evacuation time, minimum congestion risk and lowest psychological stress, are defined.

[0053] A constraint condition set is constructed, including evacuation passage quantity constraint and guidance device quantity constraint.

[0054] A Pareto optimization method is used to solve the multi-objective optimization problem.

[0055] According to the scene characteristics, the most suitable evacuation scheme is selected from the optimal solution set.

[0056] The present application provides a tunnel fire emergency linkage system based on digital twinning, which is used to execute the tunnel fire emergency linkage method based on digital twinning.

[0057] A situation awareness module is configured to acquire fire environment data through a sensor network and construct an overall situation.

[0058] A collaborative analysis module is configured to process environment and crowd characteristic information and realize multi-dimensional information fusion.

[0059] An adaptive attention module is configured to dynamically allocate attention priority according to fire development characteristics.

[0060] A behavior prediction module is configured to construct a psychological response and behavior model of the crowd under different environmental conditions.

[0061] A strategy generation module is configured to create a differentiated evacuation guidance scheme according to scene characteristics and crowd state.

[0062] A system optimization module is configured to evaluate the guiding effect and realize dynamic adjustment of parameters through virtual-real mapping.

[0063] The tunnel fire emergency linkage method based on digital twinning has the advantages that: the multi-head attention network simultaneously processes the environmental state and the crowd behavior characteristics, the system can understand the internal correlation between the environmental factors and the crowd behavior, thereby forming a comprehensive and accurate cognition of the tunnel fire scene;

[0064] Through the context-enhanced attention mechanism, the system can dynamically adjust the attention weight of different feature dimensions of the crowd according to the fire characteristics, so that the system can adaptively focus on the most critical crowd behavior characteristics in the current scene, and the blind spot problem caused by the fixed attention allocation in the traditional system is avoided;

[0065] The psychological environment interaction model enables the system to predict the evolution of crowd behavior in different fire scenes and identify potential dangerous states in advance, which solves the problem of insufficient understanding of the evolution law of crowd behavior in the fire scene in the traditional system and avoids the possible misjudgment at the critical moment;

[0066] The multi-level evacuation guiding strategy tree realizes the precise guiding strategy closely combined with the specific situation, provides differentiated guidance according to the crowd in different regions and different psychological states, and overcomes the problem that the traditional guiding strategy is out of touch with the actual situation;

[0067] By synchronizing the global situation understanding and guiding effect data to the digital twin model, the system can continuously optimize the parameters of each module based on historical data, improve the overall system performance, and realize the continuous evolution of the system;

[0068] The application of the multi-objective optimization algorithm enables the system to effectively reduce the psychological pressure of the crowd while ensuring the evacuation efficiency, and avoids the panic and stampede risk caused by ignoring the psychological factors of the crowd while pursuing the evacuation efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0069] Figure 1 is a flowchart of a tunnel fire emergency linkage method based on digital twinning in the present application;

[0070] Figure 2 is a column chart of the dynamic adjustment effect of the context-enhanced attention mechanism on the weight of different attention heads in the fire scene;

[0071] Figure 3 is a column chart of the analysis of three main psychological state indicators of the crowd in different regions;

[0072] Figure 4 is a scatter plot of the application result distribution of the multi-objective optimization algorithm;

[0073] Figure 5 is a combination chart of the real-time guiding effect evaluation comparison;

[0074] Figure 6 is a bar chart showing the performance comparison between the system and the traditional system in key performance indicators. DETAILED DESCRIPTION

[0075] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that discussions of these implementations are merely provided to enable those skilled in the art to better understand so as to be able to implement the subject matter described herein, and variations of elements discussed can be made by one skilled in the art, without departing from the scope of the present specification. Various examples can omit, substitute, or add various procedures or components as appropriate, and the methods described can also be able to be performed in an order different from that described, and that various other implementations can exist. Also, features described in relation to one example can be combined in any other example.

[0076] In at least one embodiment of the present disclosure, a tunnel fire emergency linkage method based on digital twinning is disclosed, as shown in Figure 1 The method comprises the following steps:

[0077] Step 1, collecting multi-modal data of fire environment by multiple sensors and constructing global situation understanding model;

[0078] In this step, the system collects multi-modal data of fire environment through multiple sensors (including but not limited to temperature sensor, smoke detector, infrared camera, visible light camera, gas concentration detector, etc.) deployed at different positions in the tunnel. Specifically, it includes:

[0079] Step 1.1, data collection;

[0080] Data is collected from each sensor node at regular intervals and transmitted to the central processing system through the communication network in the tunnel;

[0081] Step 1.2, data preprocessing;

[0082] The collected data is preprocessed by denoising, completing, normalizing, etc., wherein:

[0083] The temperature data is normalized to the minimum and maximum, and mapped to the range [0, 1];

[0084] The smoke concentration data is standardized and converted to Z-score;

[0085] The gas concentration data is segmented and normalized according to the danger threshold;

[0086] The image data is preprocessed by grayscale, contrast enhancement, etc.

[0087] Step 1.3, spatio-temporal alignment;

[0088] The data collected by different sensors and at different positions are aligned in time and space dimensions to form a unified data representation;

[0089] Step 1.4, feature extraction;

[0090] For data of different modalities, a special feature extraction network is used to extract corresponding feature representations;

[0091] Step 1.5, multi-modal fusion;

[0092] A multi-modal fusion network based on attention mechanism is used to fuse features of different modalities to construct a global situation representation of the tunnel fire, including fire spread range, smoke distribution, temperature distribution, and toxic gas concentration distribution.

[0093] The global situation understanding model adopts the following structural components:

[0094] Multi-modal encoder unit: contains multiple independent encoder components, each encoder is responsible for processing data of a specific modality;

[0095] Cross-modal attention unit: calculates the correlation weight matrix between different modalities;

[0096] Feature fusion unit: weights and fuses features of each modality according to attention weights;

[0097] Situation representation unit: converts fused features into physical quantity distribution representation of tunnel space.

[0098] Further, the global situation understanding model also includes a time series processing unit for capturing the dynamic evolution trend of fire situation.

[0099] Step 2, based on the global situation understanding model, a multi-head attention network is used to process environmental state and crowd behavior features to realize collaborative perception of environment and crowd information;

[0100] In this step, the system uses a multi-head attention network to simultaneously process environmental state and crowd behavior features to realize collaborative perception of environment and crowd information. The multi-head attention network includes the following structural components:

[0101] Environment feature input unit: receives and processes environmental state features in the global situation understanding model, such as fire location, smoke concentration, temperature, and oxygen content; this unit also includes an environment feature preprocessing subunit that standardizes different physical quantity environmental parameters to ensure that environmental parameters of different dimensions can be effectively compared;

[0102] Crowd feature input unit: receives and processes the behavior characteristics of the crowd, such as position distribution, moving speed, moving direction, and gathering state; this unit includes a crowd feature preprocessing subunit that normalizes position coordinates, converts moving speed into relative speed ratio, and quantifies gathering state into density index, ensuring that different types of behavior characteristics can be represented in the same feature space;

[0103] Multi-head attention calculation unit: contains parallel attention head components, each responsible for focusing on a specific environmental crowd interaction pattern, such as:

[0104] Attention head 1: focuses on the relationship between fire location and crowd distribution;

[0105] Attention head 2: focuses on the relationship between smoke diffusion and crowd moving direction;

[0106] Attention head 3: focuses on the relationship between temperature gradient and crowd moving speed;

[0107] Attention output fusion unit: fuses the outputs of attention heads through learnable weights;

[0108] Feature representation generation unit: generates a comprehensive feature representation that contains both environmental state and crowd behavior information.

[0109] Further, before processing by the multi-head attention network, the system performs feature space alignment preprocessing on environmental features and crowd features, including:

[0110] Feature dimension unification: through feature dimension reduction or dimension increase operations, the environmental features and crowd features have the same dimension;

[0111] Feature scale standardization: Z-score standardization or Min-Max normalization processing is performed on different physical dimension features;

[0112] Feature space mapping: using feature projection matrix, features of different dimensions and different physical meanings are mapped to a common hidden space, so that environmental features (physical quantities) and crowd features (behavior characteristics) can be effectively compared and calculated in the same feature space;

[0113] Feature importance balance: through feature importance scoring, balance weights are applied to environmental features and crowd features to avoid dominance of a certain type of feature in the fusion process.

[0114] Further, the multi-head attention network also includes a residual connection unit to preserve key information in the original input features and prevent information loss during attention calculation.

[0115] Further, each attention head of the multi-head attention network includes three sub-components: query transformation, key transformation, and value transformation, which correspond to the mapping relationship between environmental features and crowd features, respectively.

[0116] Step 3: Apply the context-enhanced attention mechanism based on the characteristics of the fire scene to dynamically adjust the attention weight of the crowd feature dimensions.

[0117] This step designs a context-enhanced attention mechanism to dynamically adjust the attention weight of different feature dimensions of the crowd according to the characteristics of different fire scenes. This is one of the core innovations of the present application. The context-enhanced attention mechanism includes the following execution steps:

[0118] Step 3.1: Obtain fire scene characteristic data.

[0119] The fire scene characteristic data includes factors such as the location, size, spread speed, and smoke diffusion state of the fire.

[0120] Step 3.2: Fire scene characteristic feature extraction and encoding.

[0121] The fire scene characteristics are extracted and encoded to convert them into a scene characteristic vector representation, including the following preprocessing operations:

[0122] The fire location feature is normalized in space coordinates to be within the [0, 1] interval.

[0123] The fire size feature is converted to a logarithmic scale to handle different magnitudes of fire size.

[0124] The spread speed and smoke diffusion state features are standardized.

[0125] The scene discrete features (such as fire type and occurrence area type) are one-hot encoded into numerical vectors.

[0126] Step 3.3: Attention weight calculation.

[0127] Based on the scene characteristic vector, the attention weight for different crowd feature dimensions (such as location, speed, aggregation, and panic level) is calculated.

[0128] Step 3.4: Attention weight application.

[0129] The calculated attention weight is applied to the multi-head attention network to dynamically adjust the attention degree of different crowd feature dimensions.

[0130] Step 3.5: Importance scoring.

[0131] Based on the adjusted attention allocation, the importance of the crowd behavior features is scored to highlight the most critical features for evacuation guidance in the current scene.

[0132] Further, before the weight calculation of the context-enhanced attention mechanism, the scene characteristic factors and feature vectors need to be preprocessed to ensure they can be calculated in the same feature space:

[0133] Scene characteristic standardization: feature-level standardization is performed on different scene characteristic factors (such as fire location, scale, spread speed, etc.), eliminating dimensional and scale differences;

[0134] Feature vector normalization: L2 normalization is performed on the crowd feature vector to make it lie on the unit hypersphere, ensuring the consistency of similarity calculation;

[0135] Cross-domain feature mapping: a specific cross-domain feature transformation function is designed to map physical scene characteristics and crowd behavior features to a semantically consistent representation space.

[0136] After completing the above preprocessing, the weight calculation formula of the context-enhanced attention mechanism is as follows:

[0137] ;

[0138] where represents the attention weight of the th feature ; represents the th scene characteristic factor; represents the summation symbol; represents the relevance score of the th feature to the scene characteristic ; represents the importance weight of the th scene characteristic factor; represents the total number of scene characteristic factors.

[0139] where the inner function is a relevance calculation function, defined as:

[0140] ;

[0141] where is a learnable projection matrix used to map the scene characteristic vector and the feature vector to the same semantic space for similarity calculation; is the total number of features; represents the transpose symbol; and represent the The first and the feature vector; denotes the natural exponential function; denotes the summation symbol; denotes the first scenario characteristic factor; denotes the scenario characteristic of the first feature correlation score.

[0142] Further, the context-enhanced attention mechanism further includes an adaptive threshold adjustment step for dynamically determining the threshold of feature importance to filter out truly key features.

[0143] As shown in Figure 2 , the context-enhanced attention mechanism is shown to dynamically adjust the weights of different attention heads in a fire scenario. By comparing the differences between the original weights and the adjusted weights, it can be clearly seen how the system dynamically adjusts the focus according to the characteristics of the fire, which is one of the core innovations in this patent.

[0144] Step 4, based on the adjusted attention weights, build a psychological environment interaction model to predict the evolution of crowd behavior under different fire scenarios;

[0145] This step builds a psychological environment interaction model to predict the evolution of crowd behavior under different fire scenarios. To ensure that environmental parameters (physical quantities) and psychological state parameters (subjective indicators) can be effectively integrated, the system first performs unified representation preprocessing on these heterogeneous data:

[0146] Unified representation of physical quantities and psychological indicators: design a dual-domain representation learning framework to map physical environment parameters and psychological state indicators to the same semantic space;

[0147] Multi-scale data standardization: for data of different time scales and spatial scales, use multi-scale standardization techniques to ensure the comparability of data of different sampling frequencies and coverage ranges;

[0148] Heterogeneous data fusion preprocessing: balance the influence of environmental physical parameters and psychological state parameters in the model through feature importance weighting, to avoid that a certain type of data dominates the model behavior due to different numerical ranges or units.

[0149] The psychological environment interaction model includes the following structural components:

[0150] Environment perception unit: receives the state parameters of the tunnel environment, including smoke concentration, temperature, visibility, toxic gas concentration, etc.; this unit performs standardization processing on the environmental parameters and normalizes the deviation of each parameter relative to the safety threshold, unifying the danger levels of different physical quantities to the [0, 1] interval danger index;

[0151] Psychological state modeling unit: contains a multi-layer perceptron and a state converter, used to model the panic level, herd behavior tendency, rational decision-making ability, etc. of the crowd; this unit converts crowd behavior data into quantitative psychological state indicators, encodes and quantifies qualitative observation data (such as facial expressions, body language, etc.), and maps different sources and types of psychological state information to the same feature space;

[0152] Behavior evolution prediction unit: based on a sequence prediction network, predicts the possible behavior changes of the crowd in the subsequent time period, such as changes in moving direction, speed changes, aggregation or dispersion trends, etc.

[0153] Dangerous state recognition unit: through pattern matching and threshold judgment, it recognizes possible dangerous states, such as crowd congestion, panic stampede, and stagnation in dangerous areas, etc.

[0154] Scenario deduction unit: adopts a Monte Carlo tree search structure to deduce multiple possible scenario development paths and evaluate the effects of different guiding strategies.

[0155] Further, the psychological state modeling unit in the psychological environment interaction model adopts a hierarchical structure, including a basic emotion layer, a cognitive assessment layer, and a behavior decision-making layer.

[0156] Further, the psychological environment interaction model also includes an individual difference modeling unit to represent the reaction differences of different individuals in the crowd when facing the same environmental threat.

[0157] As shown in Figure 3 , three main psychological state indicators of the crowd in different areas are displayed: panic level, herd behavior tendency, and rational decision-making ability. Through these data, the psychological state differences of the crowd in different areas can be intuitively understood, providing basic data support for the psychological environment interaction model, and helping the system to generate more accurate evacuation guiding strategies.

[0158] Step 5, according to the prediction results, generate a situation-adaptive multi-level evacuation guiding strategy tree, and provide differentiated evacuation guidance;

[0159] This step generates a situation-adaptive multi-level evacuation guiding strategy tree based on the above analysis results. The multi-level evacuation guiding strategy tree model includes the following structural components:

[0160] Root node unit: represents the starting point of the overall evacuation decision, containing global decision parameters;

[0161] Environment state branch unit: based on the fire situation, the tunnel space is divided into multiple regional nodes, such as safe zone nodes, buffer zone nodes, and dangerous zone nodes;

[0162] Psychological state branch unit: under each regional node, further branch according to the crowd psychological state, forming secondary nodes;

[0163] Guidance strategy leaf node unit: the leaf node of the tree, containing specific guidance strategies for specific environmental state and psychological state combinations, including:

[0164] Visual guidance component: such as dynamic evacuation indicator lights, escape path identification, etc.

[0165] Voice guidance component: generate voice prompts with different tones and content for crowds with different levels of panic.

[0166] Physical guidance component: such as emergency lighting, ventilation equipment control, etc.

[0167] Priority scoring unit: assign a priority score to each leaf node strategy to ensure that resources are prioritized for areas and crowds that need the most help.

[0168] Further, each leaf node in the multi-level evacuation guidance strategy tree model contains a condition triggering rule component for defining the specific conditions for strategy activation.

[0169] Further, the multi-level evacuation guidance strategy tree model also includes a strategy conflict resolution unit to handle potential conflicts between multiple strategies.

[0170] To generate the optimal evacuation guidance strategy, this step applies a multi-objective optimization algorithm to balance evacuation efficiency and crowd psychological stress. The multi-objective optimization algorithm includes the following execution steps:

[0171] Define optimization objective functions, including the shortest evacuation time objective function , the minimum congestion risk objective function , the lowest psychological stress objective function , etc.

[0172] Further, these objective functions are defined as follows:

[0173] Evacuation time objective function:

[0174] ;

[0175] Where represents the shortest evacuation time of the scheme ; is the set of all personnel; is the time required for the personnel to complete evacuation under the scheme ; represents the maximum value function; represents the evacuation scheme; denotes a conditional separator; denotes a belongs-to symbol;

[0176] Evacuation time calculation function The specific implementation is: calculating the expected evacuation time of the i-th individual from the current position to the nearest safe exit under the scheme . This function takes into account factors such as the distance from the individual's current position to the exit, the population density on the path, the walking speed, the congestion delay, and individual characteristics (such as age, mobility). In the calculation process, a grid-based path planning algorithm is used to divide the path into multiple grid cells, calculate the travel time for each cell, and accumulate the travel times of different sections to finally obtain the total evacuation time. Congestion risk objective function:

[0177]

[0178] ;

[0179] where denotes the minimum congestion risk of the scheme ; is the set of all evacuation routes; is the congestion metric of the j-th route under the scheme ; is the route weight of the j-th route; denotes the summation symbol; denotes the belongs-to symbol; Congestion metric function The specific implementation is: calculating the congestion level index of the j-th route under the scheme

[0180] . This function is calculated based on the personnel density, flow-to-capacity ratio of the route, first determines the standard capacity of the route, then calculates the ratio of the current flow to the standard capacity, and considers the uniformity of the spatial distribution of personnel in the route. When the personnel density exceeds the critical threshold, a nonlinear penalty factor is introduced to reflect the sharp growth characteristics of the congestion risk. Finally, a normalized congestion index between 0 and 1 is output. Psychological stress objective function:

[0181] where

[0182] denotes the minimum psychological stress of the scheme ;

[0183] is the stress level of the i-th individual under the scheme ; ​​​​​​For the scenario Under the scenario The uncertainty perception level of each individual; The weight coefficient representing the stress level, The weight coefficient representing the uncertainty perception level; The summation symbol; The personnel number; The belonging symbol; The set of all personnel;

[0184] Stress level calculation function The specific implementation is to evaluate the psychological stress level of each individual under the scenario Under the scenario This function comprehensively considers environmental threat factors (such as fire, temperature, smoke concentration, etc.), individual factors (such as age, coping ability, emergency situation experience, etc.) and social factors (such as group behavior, information acquisition degree, leadership instruction clarity, etc.). The calculation process uses a hierarchical weighting method, first evaluates the stress contribution of each sub-factor, then integrates through adaptive weights, and finally outputs a stress degree score of 0 to 10.

[0185] Uncertainty perception level calculation function The specific implementation is to measure the uncertainty perception level of each individual under the scenario Under the scenario This function is based on information entropy theory to calculate the completeness, accuracy and consistency of information acquisition, and to evaluate the predictability of environmental changes. The specific calculation includes the number of information acquisition channels, information update frequency, information content consistency, environmental state change rate, etc. When information is missing, contradictory or the environment changes rapidly, the uncertainty perception level increases; on the contrary, when sufficient, consistent and stable information is obtained, the uncertainty perception level decreases. Finally, an uncertainty score of 0 to 10 is output.

[0186] Construct a set of constraints, such as the number of available evacuation channels, the number of guidance devices, resource allocation constraints, etc.

[0187] Initialize the solution space to generate an initial solution set that meets the constraint conditions;

[0188] Perform a multi-objective optimization iteration process using the Pareto optimization method to continuously improve the solution set;

[0189] Get the Pareto optimal solution set, which is a set of optimal solutions that balance different objectives;

[0190] Calculate the fitness score according to the characteristics of the current scene, and select the most suitable evacuation plan from the optimal solution set.

[0191] Further, the fitness score calculation function is defined as:

[0192] ;

[0193] wherein is the fitness score of the solution under the scenario ; is the candidate solution; is the current scenario characteristic; is the dynamic weight of the th objective function under the scenario ; is the normalized result of the th objective function; denotes the summation symbol;

[0194] The objective functions include three objective functions of evacuation time, congestion risk and psychological stress, corresponding to , and , respectively;

[0195] The calculation of the dynamic weight requires preprocessing and standardization, specifically including:

[0196] Scenario characteristic vectorization: converting the scenario characteristic into a numerical feature vector, including fire size, spread speed, smoke density, crowd density and other multi-dimensional features;

[0197] Weight dynamic mapping: designing a scenario-weight mapping function to map the scenario characteristic vector to the weight space, ensuring reasonable weight distribution of the objective functions under different scenarios;

[0198] Weight normalization: normalizing the generated weight vector to ensure that the total weight is 1;

[0199] Weight smoothing mechanism: introducing a weight smoothing factor to avoid drastic fluctuations in weight caused by slight changes in scenarios, ensuring system stability.

[0200] wherein the specific implementation of the normalization function is as follows: different normalization strategies are adopted according to the characteristics of different objective functions, linear normalization is adopted for the evacuation time objective function to map it to the [0, 1] interval, piecewise function normalization is adopted for the congestion risk objective function to reflect the risk threshold effect, and S-shaped function normalization with saturation effect is adopted for the psychological stress objective function to emphasize the marginal effect of high stress state. The function first calculates the quantile of the objective function value in the historical data, then applies the corresponding nonlinear transformation according to the objective characteristics, and finally outputs the normalized value standardized to the [0, 1] interval.

[0201] Further, data preprocessing is required before calculating the fitness score for different objective function values, including:

[0202] For the evacuation time objective function value Min-Max normalization is used to map it to the [0, 1] interval;

[0203] For the congestion risk objective function value A piecewise function is used to map it, reflecting the nonlinear growth of risk at different congestion levels;

[0204] For the psychological stress objective function value Exponential conversion and normalization are performed to appropriately emphasize the adverse effects of high stress states;

[0205] Normalization function Different normalization strategies are selected according to the characteristics of different objective functions to ensure that objective functions of different dimensions and scales can be effectively weighted and compared.

[0206] Further, the multi-objective optimization algorithm uses a solution sorting step based on the dominance relationship to improve the search efficiency of the solution space.

[0207] Further, the multi-objective optimization algorithm uses an adaptive weight adjustment strategy in the iteration process to adjust the optimization weights of each objective according to the dynamic changes of the scene.

[0208] As shown in Figure 4 , the performance of different candidate solutions generated by the multi-objective optimization algorithm in terms of evacuation time and congestion risk is shown. Through this chart, the strengths and weaknesses of different solutions and their balance can be intuitively understood, helping the system find the best compromise solution among multiple objectives, reflecting the application effect of the multi-objective optimization algorithm in this patent.

[0209] Step 6, during the implementation of the evacuation guidance strategy, real-time effect evaluation is carried out, and the system overall optimization is realized through the digital twin model;

[0210] In this step, the system evaluates the real-time effect of the generated evacuation guidance strategy, and simultaneously optimizes the overall system through the digital twin model. This step includes two closely related parts: real-time parameter adjustment and digital twin system optimization.

[0211] Step 6.1, real-time guidance effect evaluation and parameter optimization;

[0212] The system evaluates the real-time effect of the generated evacuation guidance strategy, and optimizes the guidance parameters based on feedback. The real-time guidance effect evaluation and parameter optimization algorithm includes the following execution steps:

[0213] Define the guiding effect evaluation index, including evacuation speed, crowd flow smoothness, panic emotion control degree, etc.

[0214] Collect real-time data through sensor networks to monitor the actual effect after the implementation of the guiding strategy;

[0215] Compare the actual effect with the expected effect, analyze the deviation, and use the deviation function Calculate; Before calculation, standardize and preprocess different types of effect index data, including:

[0216] Normalize the evacuation speed index (m / s) to convert it into a proportion relative to the maximum possible evacuation speed;

[0217] Standardize the crowd flow smoothness (dimensionless) so that it is distributed in the [0, 1] interval;

[0218] Convert the panic emotion control degree (based on questionnaire scores) to a standard score;

[0219] Use specific scaling factors for indicators with different units and magnitudes to make comparisons meaningful;

[0220] Further, the deviation function is defined as:

[0221] ;

[0222] Where is the deviation value; is the measured value of the th effect index (after standardization), is the expected value of the th effect index (after the same standardization), is the importance weight of the th effect index, is the total number of effect indexes; denotes the summation symbol;

[0223] According to the deviation analysis results, generate parameter adjustment amount, and adaptively adjust the guiding strategy parameters;

[0224] Apply the adjusted parameters to the guiding device to update the start time, signal strength, guiding information content, etc.

[0225] Further, the real-time guiding effect evaluation and parameter optimization algorithm also includes a parameter boundary check step to ensure that the adjusted parameters are within the effective range.

[0226] Further, the real-time guiding effect evaluation and parameter optimization algorithm uses online learning to gradually optimize the parameter adjustment strategy and improve the adjustment effect.

[0227] As shown in Figure 5 , the key indicators of the system real-time guidance effect evaluation are displayed, the left axis represents the evacuation speed of different areas, and the right axis represents the comprehensive indicator score (the weighted average of flow smoothness and panic control). By comparing the indicator performance of different areas, the effectiveness of the guidance strategy can be evaluated, providing a basis for parameter optimization, and reflecting the role of real-time guidance effect evaluation and parameter optimization mechanism in this patent.

[0228] Step 6.2, digital twin model synchronization and system optimization;

[0229] This link synchronizes global situation understanding and guidance effect data to the digital twin model, supports overall system optimization, and forms a closed-loop feedback mechanism of real-time adjustment and long-term optimization. The digital twin model includes the following structural components:

[0230] Data synchronization unit: responsible for receiving and processing real-time acquired fire situation data, crowd behavior data, and guidance effect data;

[0231] Physical-virtual data conversion preprocessing unit: necessary data preprocessing and format conversion before mapping physical world data to virtual model, including:

[0232] Data format unification: convert physical world data of different sources and formats into standard data format acceptable by the digital twin model;

[0233] Space-time reference system conversion: establish the coordinate mapping relationship between physical space and virtual space to ensure accurate spatial position data correspondence;

[0234] Data precision matching: according to the precision requirements of the virtual model, perform upsampling or downsampling processing on the physical world data;

[0235] Abnormal value detection and processing: identify and process abnormal values, missing values and noise in physical sensor data to ensure data quality;

[0236] Data semantic enhancement: add semantic labels and context information to the original data to improve the understanding ability of the virtual model to the data.

[0237] Virtual mapping unit: map the preprocessed physical entity information to the corresponding entity in the virtual space;

[0238] Three-dimensional visualization unit: visually display fire situation, crowd distribution, evacuation path, etc. through graphics rendering technology;

[0239] Historical data storage unit: save the historical data of system operation in time series form;

[0240] Parameter optimization unit: based on historical data analysis, the parameter optimization suggestions of each module of the system are proposed;

[0241] Simulation verification unit: through simulation experiment in digital space, the performance of the optimized system is verified.

[0242] Further, the digital twin model further includes a state prediction unit, which can predict the future evolution trend of the system based on the current state.

[0243] Further, the digital twin model includes a physical-virtual bidirectional interaction unit, which supports bidirectional information flow and control between the physical world and the virtual world.

[0244] As shown in Figure 6 the differences between the system and the traditional system in key performance indicators are shown, including collaborative perception ability and evacuation guidance effect. Through intuitive comparison, it can be seen that the system is superior to the traditional system in various indicators, verifying the effectiveness and advancement of the core technical scheme in the patent.

[0245] A tunnel fire emergency linkage system based on digital twin, for executing the tunnel fire emergency linkage method based on digital twin, comprising:

[0246] A situational awareness module for obtaining fire environment data through a sensor network and constructing an overall situation;

[0247] A collaborative analysis module for processing environmental and crowd characteristic information and realizing multi-dimensional information fusion;

[0248] An adaptive attention module for dynamically allocating attention priority according to fire development characteristics;

[0249] A behavior prediction module for constructing psychological response and behavior models of crowds under different environmental conditions;

[0250] A strategy generation module for creating differentiated evacuation guidance schemes according to scene characteristics and crowd state;

[0251] A system optimization module for evaluating guidance effect and realizing dynamic adjustment of parameters through virtual-real mapping.

[0252] Here, the embodiment provides an application example:

[0253] The application scenario of the present embodiment is a certain city's two-way six-lane urban tunnel, with a total length of 3.5 kilometers and a single-way three-lane, with an average daily traffic volume of about 80,000 vehicles. The tunnel is equipped with a complete sensor network, including temperature sensors (140 groups), smoke detectors (280 groups), infrared and visible light cameras (210 groups), toxic gas detectors (70 groups), LED dynamic evacuation indicator lights (700 groups), and a voice broadcasting system (70 groups). On June 15, 2023, a chemical truck collision and fire accident occurred in the tunnel, and there were about 420 people in the tunnel who needed to be evacuated urgently. The system was applied in this emergency and the technical effect was verified.

[0254] After the accident, various sensors in the tunnel quickly collected environmental data and pre-processed them (temperature data normalization, smoke concentration standardization, image data feature extraction, etc.). The fire situation assessment results generated by the system after processing are shown in Table 1:

[0255] Table 1: Fire situation assessment results

[0256]

[0257] At the same time, the system collects crowd distribution and behavior data through cameras and sensors, and identifies and analyzes the crowd state in different areas. The multi-head attention network processes both environmental state and crowd behavior features, focusing on different environmental-crowd interaction patterns. The different regional crowd states identified by the system are shown in Table 2:

[0258] Table 2: Different regional crowd state identification results

[0259]

[0260] The context-enhanced attention mechanism dynamically adjusts the attention weights of different feature dimensions of the crowd according to the current fire scene characteristics. The system identifies that the fire spreads rapidly, the smoke concentration is extremely high, and the crowd is in panic and some reverse movement, so it automatically increases the attention weights of smoke diffusion-movement direction and toxic gas concentration-crowd density, ensuring that these key factors are focused on, and the originally evenly distributed weights (0.25 each) are dynamically adjusted to a more suitable distribution for the current scene (smoke diffusion-movement direction: 0.35, toxic gas concentration-crowd density: 0.30, fire location-crowd distribution: 0.20, temperature gradient-movement speed: 0.15).

[0261] The psychological environment interaction model predicts the evolution of crowd behavior in the fire scene based on environmental state and crowd behavior data. The system models the psychological state and risk assessment of different areas, and the psychological state modeling and evacuation guidance strategy are shown in Table 3:

[0262] Table 3: Psychological state modeling and evacuation guidance strategy

[0263]

[0264] Based on the fire situation assessment and the psychological state of the crowd, the system constructs a multi-level evacuation guidance strategy tree. The strategy tree takes environmental state and psychological state as branch conditions, generating targeted guidance strategies. The system assigns guidance priorities to different areas, ensuring that resources are prioritized for the most dangerous areas and the most in need of help. For example, the 0-100 meter area (extremely high panic area) uses short and clear instructions and a 2Hz frequency flashing red arrow, while reducing the alarm volume to avoid increasing panic; while the area far from the accident uses a more conventional guidance method.

[0265] The system evaluates the effect of the guidance strategy in real time during the evacuation process and optimizes the guidance parameters based on feedback. Through the sensor network, the system continuously monitors key indicators such as evacuation speed, crowd flow smoothness, and panic control, calculates the deviation between actual and expected effects using a deviation function, and generates appropriate parameter adjustment schemes. The evaluation and adjustment of the system are shown in Table 4:

[0266] Table 4: Evaluation of guidance strategy effect and parameter adjustment

[0267]

[0268] During the evacuation process, the system synchronizes global situation understanding and guidance effect data to the digital twin model, supporting overall system optimization. The digital twin model constructs a virtual mapping of the physical tunnel, realizing physical-virtual bidirectional interaction. Through this model, the system can map physical world data to virtual space, perform visual display, future state prediction, and parameter optimization. The prediction and optimization results of the digital twin model at key time points are shown in Table 5:

[0269] Table 5: Prediction and optimization results of digital twin model

[0270]

[0271] During the evacuation process, the system applies a multi-objective optimization algorithm to balance evacuation efficiency and crowd psychological pressure. This algorithm optimizes three key objectives: shortest evacuation time, minimum congestion risk, and lowest psychological pressure, considering constraints such as the number of channels and guidance device resources. The system uses the Pareto optimization method to generate multiple candidate solutions and calculates fitness scores based on current scene characteristics to select the most suitable evacuation scheme. The evaluation results of the candidate solutions generated by the system are shown in Table 6:

[0272] Table 6: Evaluation results of multi-objective optimization schemes

[0273]

[0274] The system selects scheme A as the final implementation scheme, although its evacuation time is slightly longer than that of scheme B, but considering the congestion risk and psychological pressure, the overall fitness score is the highest, which can effectively reduce the psychological pressure of the crowd while ensuring the evacuation efficiency, and avoid the congestion and panic risk caused by simply pursuing the evacuation time and ignoring other factors.

[0275] By comparing the performance of the technical solution and the traditional tunnel fire emergency system under the same conditions, the two core technical effects are verified: the cooperative perception and analysis ability of fire situation and crowd psychological state; and the differentiated precise evacuation guidance effect.

[0276] The system performance comprehensive comparison results are shown in Table 7:

[0277] Table 7: System performance comprehensive comparison results

[0278]

[0279] In terms of cooperative perception ability, the system improves the recognition rate of the correlation between environmental factors and crowd behavior, and increases the psychological state assessment accuracy by 41.3%, so that the system can more accurately understand the complex fire scene. In terms of evacuation guidance effect, the system shortens the evacuation completion time by 31.1%, while reducing the local congestion occurrence rate and the proportion of reverse crowd, and improving the response degree of the guidance strategy.

[0280] The context-enhanced attention mechanism, as the core innovation point of the system, effectively bridges the gap between fire situation understanding and crowd psychological state analysis, providing a solid foundation for precise guidance. By dynamically adjusting the attention weight of different environmental and crowd characteristics, the system can develop targeted evacuation strategies, improving the efficiency and safety of personnel evacuation and reducing the potential risk of casualties in accidents.

[0281] The above describes the embodiments of the present application, but the embodiments are not limited to the specific implementation described above, which is only illustrative and not limiting, and those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection of the embodiments.

Claims

1. A tunnel fire emergency linkage method based on digital twinning, characterized in that, The application relates to a fire evacuation guidance system based on multi-modal data fusion and multi-head attention network. The system includes: Collecting multi-modal data of fire environment through various sensors and constructing a global situation understanding model; Based on the global situation understanding model, a multi-head attention network is used to process environmental state and crowd behavior characteristics, realizing the collaborative perception of environment and crowd information; Combined with the characteristics of fire scenes, a context-enhanced attention mechanism is applied to dynamically adjust the attention weight of crowd feature dimensions; Based on the adjusted attention weight, a psychological environment interaction model is constructed to predict the evolution of crowd behavior under different fire scenes; According to the prediction results, a multi-level evacuation guidance strategy tree is generated to provide differentiated evacuation guidance; 2. The tunnel fire emergency linkage method based on digital twinning according to claim 1, characterized in that, During the implementation of the evacuation guidance strategy, real-time effect evaluation is carried out, and the overall system optimization is realized through a digital twin model. The global situation understanding model includes: A multi-modal encoder unit for processing different modal sensor data; A cross-modal attention unit for calculating the correlation weight matrix between different modalities; A feature fusion unit for weighting and fusing the features of each modality according to the attention weight; 3. The tunnel fire emergency linkage method based on digital twinning according to claim 1, characterized in that, A situation representation unit for converting the fused features into a physical quantity distribution representation of the tunnel space. The multi-head attention network includes: An environmental feature input unit for receiving and processing environmental state features in the global situation understanding model; The multi-head attention computing unit comprises An attention head is responsible for focusing on different environmental crowd interaction modes respectively. An attention output fusion unit is configured to weight and fuse the outputs of the attention heads.

4. The tunnel fire emergency linkage method based on digital twinning according to claim 1, characterized in that, A crowd feature input unit for receiving and processing crowd behavior characteristics; The psychological environment interaction model includes: An environmental perception unit for receiving environmental state parameters; A psychological state modeling unit for modeling the psychological state of the crowd; A behavior evolution prediction unit for predicting the change of crowd behavior; 5. The tunnel fire emergency linkage method based on digital twinning according to claim 1, characterized in that, A dangerous state recognition unit for recognizing dangerous states. The multi-level evacuation guidance strategy tree includes: A root node unit representing the starting point of the overall evacuation decision; An environmental state branch unit for dividing the tunnel space into different regional nodes; A psychological state branch unit for further branching according to the psychological state of the crowd; A guidance strategy leaf node unit containing specific guidance strategies for specific combinations of environmental state and psychological state; 6. The tunnel fire emergency linkage method based on digital twinning according to claim 1, characterized in that, A priority scoring unit for assigning priorities to each guidance strategy. The context-enhanced attention mechanism includes: Obtaining fire scene characteristic data, including the location, size, spread speed, and smoke diffusion state of the fire; Feature extraction and encoding of fire scene characteristics, converting them into a scene characteristic vector representation; According to the scene characteristic vector, calculate the attention weight for different crowd feature dimensions; Apply the calculated attention weight to the multi-head attention network to dynamically adjust the attention degree of different crowd feature dimensions; 7. The tunnel fire emergency linkage method based on digital twinning according to claim 1, characterized in that, Based on the adjusted attention distribution, score the importance of crowd behavior characteristics, highlighting the most critical features for evacuation guidance in the current scene. The real-time effect evaluation includes: Define guidance effect evaluation indicators, including evacuation speed, crowd flow smoothness, and panic emotion control degree; Collect real-time data through a sensor network to monitor the execution effect of the guidance strategy; Use a bias function to calculate the deviation between the measured effect and the expected effect; 8. The tunnel fire emergency linkage method based on digital twinning according to claim 1, characterized in that, Generate parameter adjustment amount based on the deviation analysis result. The digital twin model includes: A data synchronization unit for receiving and processing real-time situation data; A virtual mapping unit is configured to map physical entity information to a virtual space; A three-dimensional visualization unit is configured to display fire situation and crowd distribution; A historical data storage unit is configured to save historical data of system operation; A parameter optimization unit is configured to optimize parameters of each module of the system.

9. The tunnel fire emergency linkage method based on digital twinning according to claim 1, characterized in that, The generated context-adaptive multi-level evacuation guidance strategy tree uses a multi-objective optimization algorithm, which includes: Defining three optimization objectives of shortest evacuation time, minimum congestion risk and lowest psychological stress; Building a constraint condition set, including evacuation passage number constraint and guidance device number constraint; Using a Pareto optimization method to solve the multi-objective optimization problem; According to the characteristics of the scene, the most suitable evacuation scheme is selected from the optimal solution set.

10. A tunnel fire emergency linkage system based on digital twinning, characterized in that, A tunnel fire emergency linkage method based on digital twinning for executing any one of claims 1-9, comprising: A situational awareness module is configured to obtain fire environment data through a sensor network and build an overall situation; A collaborative analysis module is configured to process environmental and crowd characteristic information and realize multi-dimensional information fusion; An adaptive attention module is configured to dynamically allocate attention priority according to the characteristics of fire development; A behavior prediction module is configured to build a psychological response and behavior model of the crowd under different environmental conditions; A strategy generation module is configured to create a differentiated evacuation guidance scheme according to the characteristics of the scene and the state of the crowd; A system optimization module is configured to evaluate the guidance effect and realize dynamic adjustment of parameters through virtual-real mapping.

Images