Intelligent fire safety early warning system for expressway tunnel

Abstract

The invention relates to the technical field of highway tunnel safety monitoring, and discloses a highway tunnel intelligent fire safety early warning system, which comprises a data acquisition and preprocessing module used for acquiring video monitoring data, environment sensor data and vehicle positioning data in a tunnel in real time; the dynamic risk assessment module is used for constructing a risk field model based on vehicle trajectory and smoke diffusion space-time coupling; the emergency resource elastic adaptation module is used for automatically adjusting the induction screen and the fire fighting equipment according to the risk intensity; according to the method, the limitation of traditional threshold alarm is avoided, fire early warning nodes are greatly advanced through multi-source data dynamic fusion, long-term stability of a monitoring benchmark is guaranteed by utilizing an optical flow compensation and pollution self-repairing mechanism, and the self-adaptive migration of a calculation task is realized by utilizing the optical flow compensation and pollution self-repairing mechanism. And the initiative and reliability of tunnel safety prevention and control are obviously improved.

Description

Technical Field

[0001] The present invention relates to an intelligent fire safety early warning system for highway tunnels, belonging to the technical field of highway tunnel safety monitoring. Background Art

[0002] As a semi-closed traffic facility, highway tunnels require real-time perception of environmental parameters, vehicle status, and emergency resource distribution for fire prevention and control. Traditional systems usually adopt a threshold-triggered alarm mechanism. For example, when the temperature or smoke concentration exceeds a preset threshold, a fire response is initiated. Although such methods can handle open fire scenarios, they have significant deficiencies in early smoldering recognition, dynamic resource scheduling, and complex data fusion:

[0003] 1. Existing systems rely on independent judgment of single-dimensional parameters such as temperature and smoke concentration, without considering the spatio-temporal misalignment between the smoke diffusion path caused by air flow disturbance in the tunnel and the vehicle trajectory. For example, when video monitoring captures the smoke texture, traditional methods are prone to false alarms or missed alarms due to the lack of coupling analysis with the gas diffusion path and being affected by light refraction or local air flow interference. To solve this problem, the industry usually increases the sensor density, but the problem of multi-source data isolation is exacerbated, and the timestamp alignment error of different modality data is further amplified, making it difficult to construct a global risk field model.

[0004] 2. Current fire linkage is mostly based on a preset static path planning, without integrating the dynamic game relationship between real-time traffic flow and fire spread speed. For example, when a heavy truck is blocked, the fire truck cannot pass through the preset route, and existing systems lack the collaborative scheduling ability of on-vehicle mobile fire extinguishing devices and fixed fire hydrants. Although some solutions introduce path optimization algorithms, they do not consider the physical constraint that the on-vehicle water tank capacity decays with the driving speed, resulting in an imbalance between the fire extinguishing agent delivery efficiency and path feasibility.

[0005] 3. Existing distributed systems mostly use independent nodes to process local data, without realizing cross-node model parameter migration and task takeover. For example, when a certain node has a monitoring benchmark drift due to hardware failure or pollution, neighboring nodes cannot quickly reconstruct the data repair matrix due to the lack of a topology awareness mechanism, resulting in the interruption of local risk field deduction; the industry tries to improve reliability through redundant deployment, but the problem of fragmented computing resources is prominent and cannot adapt to the characteristics of long-distance linear distribution of tunnels.

[0006] Therefore, how to achieve spatio-temporal coupling analysis of multi-source data, dynamic resource elastic adaptation, and edge node collaborative fault tolerance has become the technical problem to be solved by the present invention. Summary of the Invention

[0007] The present invention provides an intelligent fire safety early warning system for highway tunnels, whose main purpose is to solve the problems of early warning lag, emergency resource misallocation, and insufficient system robustness.

[0008] To achieve the above object, a smart fire safety warning system for highway tunnels provided by the present invention includes:

[0009] A data acquisition and preprocessing module, which is used to collect multi-source heterogeneous data in the tunnel in real time. The multi-source heterogeneous data at least includes video surveillance data, environmental sensor data, and vehicle positioning data; and extract vehicle movement trajectories and smoke texture features from the video surveillance data, perform spatio-temporal interpolation compensation on the environmental sensor data, and perform cross-modal early smoldering feature recognition on the video pixel change rate ΔP in the video surveillance data and the carbon monoxide concentration change gradient in the environmental sensor data. When is satisfied, it is determined as early smoldering;

[0010] A dynamic risk assessment module, which is communicatively connected to the data acquisition and preprocessing module, and is used to construct a fire risk field model based on the spatio-temporal coupling of multi-source heterogeneous data. Specifically, it includes: generating a risk propagation vector field based on vehicle movement trajectories and smoke texture features, combining static parameters such as the positions of fire hydrants and the states of smoke exhaust ports in the tunnel and the real-time vehicle distribution to generate a dynamic accessibility heat map, and establishing a game model between the fire spread speed and the traffic flow speed to predict the optimal rescue time window;

[0011] An emergency resource elastic adaptation module, which is communicatively connected to the dynamic risk assessment module, and is used to automatically switch the display strategy of the induction screen according to the intensity gradient of the fire risk field model, and form a negative feedback closed loop for the pressure regulation of the fire pump and the rotation speed of the smoke exhaust fan to suppress the backfire of the flue gas;

[0012] Edge computing nodes, which are distributed in the tunnel, are used to execute the computing tasks of the data acquisition and preprocessing module and the dynamic risk assessment module, and realize the adaptive transfer learning of model parameters through topology awareness, and seamlessly take over the computing tasks of adjacent nodes in case of single-point failure.

[0013] Preferably, the data acquisition and preprocessing module further includes an optical flow compensation mechanism, specifically: using the optical flow field of the fixed light source in the tunnel sidewall monitoring video to invert the real-time air flow velocity field, and dynamically correcting the spatio-temporal mapping relationship between the smoke diffusion path and the vehicle trajectory based on the real-time air flow velocity field to compensate for the non-linear spatio-temporal distortion caused by air flow disturbance.

[0014] Preferably, the data acquisition and preprocessing module further includes a local pollution perception and image self-repair mechanism, specifically including: identifying the pollution characteristics on the surface of the fixed light source based on a lightweight model to generate a pollution area mask; using the spatial symmetry of adjacent fixed light sources and the temporal illumination consistency to construct a multi-frame reference repair matrix for the pollution area; and when the pollution area ratio S polluted / S total > θ2 and lasts for a preset time Tpolluted When the humidity data in the tunnel meets a certain condition, the repair algorithm is activated, and non-emergency repairs are suppressed according to the humidity data in the tunnel, where S polluted is the area of the polluted area, and S total is the total area of the light sources.

[0015] Preferably, the local pollution perception and image self-repair mechanism further includes: when it is recognized that the polluted area overlaps with the predicted path of the vehicle trajectory, this area is preferentially repaired; and the sampling frequency of the repair area is dynamically adjusted based on the predicted trajectory speed; and the repaired optical flow data is fed back for optimizing the vehicle trajectory prediction model.

[0016] Preferably, the game model of the fire spread speed and the traffic flow speed established by the dynamic risk assessment module also considers the impact of heavy truck jams on the fire truck's travel route, and combines the collaborative scheduling capabilities of on-vehicle mobile fire extinguishing devices and fixed fire hydrants to generate an optimal fire extinguishing resource scheduling plan.

[0017] Preferably, when generating the optimal fire extinguishing resource scheduling plan, the dynamic risk assessment module introduces an on-vehicle water tank capacity attenuation factor and calculates the Pareto front solution set of the optimal path and injection pressure in real time.

[0018] Preferably, when forming a negative feedback closed loop of the fire pump pressure regulation and the exhaust fan speed, the emergency resource elastic adaptation module embeds a second derivative feedback term of the fire water pressure sensor in the exhaust fan control loop to dynamically suppress the air turbulence caused by sudden changes in water pressure.

[0019] Preferably, when generating the risk propagation vector field, the dynamic risk assessment module performs microsecond-level temporal alignment of the vehicle movement trajectories in the video stream and the smoke diffusion direction to compensate for the temporal misalignment between the smoke diffusion path and the video movement trajectory caused by tunnel air flow disturbances.

[0020] Preferably, when calculating the time-space game equilibrium point of the evacuation path and the optimal travel route of the fire truck, the dynamic risk assessment module introduces a fire truck turning radius constraint to generate an optimal path with continuous curvature.

[0021] Preferably, the edge computing nodes broadcast their respective local risk assessment results and resource status information periodically to achieve collaborative perception of the overall tunnel risk situation and distributed collaborative scheduling of emergency resources.

[0022] Compared with the problems in the background art, the beneficial effects of the present invention are:

[0023] 1. Through the dynamic coupling analysis of video pixel change characteristics and environmental parameter gradients, the system can capture the implicit correlation between trace heat release and gas diffusion that traditional single sensors cannot identify. When the spatiotemporal propagation path of the smoke texture forms a specific phase difference with the concentration field, the system automatically triggers the early intervention strategy to move the fire warning node forward to the smoldering stage. This mechanism effectively compensates for the monitoring distortion caused by airflow disturbances in the tunnel and realizes the self-consistency verification of multi-physical field data.

[0024] 2. By integrating vehicle trajectories, fire-fighting facility status and smoke diffusion vectors, a dynamic game space for emergency resource supply and demand is constructed. The system automatically generates a fire truck's travel path with continuous curvature and an optimal pressure-flow solution set by deducing the speed game relationship between fire spread and traffic flow in real time. This dynamic mapping mechanism not only avoids the path conflict risk of fixed plans, but also effectively suppresses the turbulent interference caused by high-pressure water mist injection through closed-loop feedback of water pump pressure and smoke exhaust speed.

[0025] 3. Based on the spatial symmetry of the LED array, the distributed nodes trigger cross-node data repair and model migration in the event of local pollution or hardware failure. By reusing the optical flow information of the traffic camera to invert environmental parameters, the system achieves dynamic load balancing of computing tasks while maintaining the integrity of the benchmark data. This self-healing mechanism ensures the continuity of the risk deduction process under a single point failure, forming a robust control network with spatiotemporal redundancy characteristics.

[0026] 4. In response to the monitoring benchmark drift caused by tunnel wall pollution, optical compensation is achieved without hardware modification through multi-frame reference repair of adjacent light sources and dynamic triggering mechanism of humidity perception. The repaired video data reversely optimizes the vehicle trajectory prediction model to form a closed-loop enhancement circuit of perception-decision-making. This mechanism not only maintains the long-term stability of optical flow field inversion, but also simultaneously improves the accuracy of derivative functions such as license plate recognition. In the scenario of heavy vehicle congestion, the system constructs a collaborative injection network of mobile fire extinguishing devices and fixed fire hydrants through the joint optimization of the on-board water tank capacity attenuation model and path curvature constraints. This dynamic adaptation mechanism avoids the static resource allocation limitations of traditional plans, and achieves the optimal balance between the fire extinguishing agent delivery efficiency and the maneuverability of fire trucks while maintaining the water flow coverage density. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a structural block diagram of the tunnel intelligent fire safety early warning system of the present invention;

[0028] Figure 2 It is a timing diagram of dynamic collaborative response based on risk parameters in the tunnel intelligent fire protection system of the present invention;

[0029] Figure 3 It is a schematic diagram of the node collaboration and fault tolerance mechanism based on the edge computing network of the present invention.

[0030] The implementation, functional features, and advantages of the present invention will be further described in conjunction with embodiments and with reference to the accompanying drawings. Specific embodiments

[0031] It should be understood that the specific embodiments described herein are merely used to explain the present invention and are not used to limit the present invention.

[0032] An embodiment of the present application provides an intelligent fire safety early warning system for highway tunnels, and the system includes:

[0033] A data acquisition and preprocessing module, configured to collect multi-source heterogeneous data in the tunnel in real time. The multi-source heterogeneous data at least includes video surveillance data, environmental sensor data, and vehicle positioning data; and extract vehicle movement trajectories and smoke texture features from the video surveillance data, perform spatio-temporal interpolation compensation on the environmental sensor data, and perform cross-modal early smoldering feature recognition on the video pixel change rate ΔP in the video surveillance data and the carbon monoxide concentration change gradient in the environmental sensor data. When is satisfied, it is determined as early smoldering.

[0034] A dynamic risk assessment module, communicatively connected to the data acquisition and preprocessing module, configured to construct a fire risk field model based on the spatio-temporal coupling of multi-source heterogeneous data, specifically including: generating a risk propagation vector field based on vehicle movement trajectories and smoke texture features, and combining static parameters such as the positions of fire hydrants and the states of smoke exhaust outlets in the tunnel and the real-time vehicle distribution to generate a dynamic reachability heat map, and establishing a game model between the fire spread speed and the traffic flow speed to predict the optimal rescue time window.

[0035] An emergency resource elastic adaptation module, communicatively connected to the dynamic risk assessment module, configured to automatically switch the display strategy of the induction screen according to the intensity gradient of the fire risk field model, and form a negative feedback closed loop for the pressure regulation of the fire pump and the rotational speed of the smoke exhaust fan to suppress the backfire of the flue gas.

[0036] Edge computing nodes, distributed in the tunnel, configured to execute the computing tasks of the data acquisition and preprocessing module and the dynamic risk assessment module, and realize the adaptive transfer learning of model parameters through topology awareness, and seamlessly take over the computing tasks of adjacent nodes in case of a single point of failure.

[0037] Preferably, the data acquisition and preprocessing module further includes an optical flow compensation mechanism, specifically: using the optical flow field of a fixed light source in the tunnel sidewall monitoring video to invert the real-time air flow velocity field, and dynamically correcting the spatio-temporal mapping relationship between the smoke diffusion path and the vehicle trajectory based on the real-time air flow velocity field to compensate for the non-linear spatio-temporal distortion caused by air flow disturbance.

[0038] Preferably, the data acquisition and preprocessing module further includes a local pollution perception and image self-repair mechanism, specifically including: identifying pollution features on the surface of the fixed light source based on a lightweight model to generate a pollution area mask; constructing a multi-frame reference repair matrix for the pollution area by utilizing the spatial symmetry of adjacent fixed light sources and the temporal illumination consistency; and when the ratio S polluted / S total > θ2 and lasts for a preset time T polluted , activating the repair algorithm and suppressing non-emergency repairs according to the humidity data in the tunnel, where S polluted is the area of the pollution area, and S total is the total area of the light source.

[0039] Preferably, the local pollution perception and image self-repair mechanism further includes: when it is identified that the pollution area overlaps with the predicted path of the vehicle, giving priority to repairing this area; dynamically adjusting the sampling frequency of the repair area based on the predicted trajectory speed; and the repaired optical flow data is fed back to optimize the vehicle trajectory prediction model.

[0040] Preferably, the game model of the fire spread speed and the traffic flow speed established by the dynamic risk assessment module also considers the impact of heavy truck blockages on the fire truck's travel route, and combines the coordinated scheduling capabilities of on-vehicle mobile fire extinguishing devices and fixed fire hydrants to generate an optimal fire extinguishing resource scheduling plan.

[0041] Preferably, when generating the optimal fire extinguishing resource scheduling plan, the dynamic risk assessment module introduces a decay factor for the on-vehicle water tank capacity and calculates the Pareto front solution set of the optimal path and injection pressure in real time.

[0042] Preferably, when forming a negative feedback closed loop for the fire pump pressure regulation and the smoke exhaust fan speed, the emergency resource elastic adaptation module embeds a second-order derivative feedback term of the fire water pressure sensor in the smoke exhaust fan control loop to dynamically suppress the air turbulence caused by sudden changes in water pressure.

[0043] Preferably, when generating the risk propagation vector field, the dynamic risk assessment module performs microsecond-level temporal alignment of the vehicle movement trajectories in the video stream and the smoke diffusion direction to compensate for the temporal misalignment between the smoke diffusion path and the video movement trajectory caused by tunnel air flow disturbances.

[0044] Preferably, when calculating the time-space game equilibrium point of the evacuation path and the optimal travel route of the fire truck, the dynamic risk assessment module introduces a turning radius constraint for the fire truck to generate an optimal path with continuous curvature.

[0045] Preferably, the edge computing nodes broadcast their respective local risk assessment results and resource status information periodically to achieve collaborative perception of the overall risk situation in the tunnel and distributed collaborative scheduling of emergency resources.

[0046] Example 1: In the scenario of a two-way highway tunnel, for multiple key aspects such as early smoldering recognition, risk field construction, resource dynamic scheduling, and edge fault tolerance, the system is deployed in a distributed edge node located at about one-third of the position from the tunnel entrance. Relying on local high-definition cameras, environmental sensor arrays, and edge computing modules, it realizes the rapid recognition of early fire signs and resource adaptive response. In terms of early smoldering recognition, the system adopts a multi-source heterogeneous data collaborative analysis mechanism. Specifically, for video surveillance data, through background modeling and foreground separation methods, the local temporal brightness change rate is extracted to form a real-time video pixel change rate sequence. At the same time, the local carbon monoxide concentration data obtained by the supporting sensors is used to obtain a carbon monoxide concentration change gradient field with strong spatial consistency through three-dimensional spline interpolation and gradient regression methods. The system performs a point-to-point multiplication operation on the real-time video pixel change rate and the carbon monoxide concentration change gradient through a set combination criterion to evaluate the synchrony of trace smoldering signs. When this product value continuously exceeds the set threshold, the system triggers an early fire response. The set threshold refers to the cross-statistical distribution of the gas release amount and visible light disturbance level in the initial stage of smoldering in similar tunnel cases in the past three years, and is set as a quantile value that balances sensitivity and false alarm rate, ensuring that the system not only has the ability of early response but also can effectively suppress false alarms.

[0047] For the coupled modeling of smoke diffusion and vehicle trajectories, the system synchronously extracts vehicle movement trajectories and changes in the smoke texture boundary at the video pixel level. Combining with the airflow velocity field obtained by optical flow inversion on the tunnel sidewall, a risk propagation vector field containing multiple moments and multiple spatial levels is constructed in each processing cycle. The vector field uses the predicted vehicle path as the reference coordinate system, the smoke edge expansion trend as the vector direction, and introduces an airflow perturbation correction factor to dynamically compensate for nonlinear perturbations, thereby generating a risk migration path map with temporal consistency and spatial accessibility. On this basis, the system further constructs a dynamic accessibility heat map to describe the spatial distribution of the degree of influence of different vehicles by the fire source from the current moment to the future estimated time window. The heat map is generated by superimposing the risk propagation vector and the real-time vehicle position, and a weighted accessibility index is constructed by combining static and semi-static parameters such as the position of fire hydrants, the status of smoke exhaust outlets, and traffic density for subsequent resource scheduling judgment. At the same time, to achieve efficient scheduling and adaptation of fire resources, the system introduces a game model of traffic flow and fire spread speed. In the model, the traffic flow speed is calculated from the average instantaneous speed of adjacent vehicles, and the fire spread speed is estimated based on the smoke texture expansion speed and the local temperature rise speed. The two act as opposing sides in the model, and the goal is to solve for the shortest response time window that a fire truck can achieve on different paths. The game model is solved iteratively based on differential dynamic programming. In each iteration, the vehicle's passable time period and the expansion boundary of the fire source influence area are updated, and finally, the shortest response path set is obtained. In the path feasibility analysis, a water tank capacity attenuation factor of the on-vehicle fire extinguishing device is introduced, that is, as the driving distance increases, the fire extinguishing agent carrying capacity decreases linearly. Therefore, the system normalizes the attenuation curve in the path planning to form a time-dose composite optimization objective and selects the optimal Pareto solution set as the scheduling recommendation.

[0048] In terms of the reliability of environmental perception, the system models and identifies the surface contamination of fixed light sources. First, by periodically collecting light source image frames, brightness distribution and edge sharpness indicators are extracted, and combined with a lightweight contamination recognition module based on a convolutional neural network to judge the degree of surface contamination of the light source and generate a contamination mask image. When the contaminated mask area overlaps with the predicted vehicle trajectory path, the system preferentially activates the image restoration module. This module uses the symmetry of adjacent light sources in spatial position and the illumination consistency between historical frames to construct a set of multi-frame reference images, and realizes the restoration of the contaminated area through weighted averaging and image reconstruction algorithms. If the contaminated area accounts for more than 5% of the total light-emitting area and the duration exceeds a set threshold (this threshold is statistically obtained based on the daily contamination evolution rate of the monitoring point and is set to more than 5 consecutive minutes), the system automatically triggers the restoration action, and the restored image is fed back to the vehicle trajectory prediction module to update its prediction model parameters to prevent the prediction path from deviating due to perception errors.

[0049] In the linkage control of the fire pump pressure and the speed of the smoke exhaust fan, the system constructs a negative feedback closed-loop. The specific mechanism is as follows: The data of the fire water pressure sensor is monitored in real time, and its second derivative is extracted to judge the trend of pressure mutation. If a pressure increase trend is detected, the system immediately reduces the speed of the smoke exhaust fan to suppress the reverse push phenomenon of local air turbulence caused by the high-pressure water column. The weight of this feedback term is derived from the hydrodynamic model of the tunnel cross-section. By adjusting the feedback sensitivity coefficient, the control system can not only respond quickly but also avoid frequent oscillations, thus realizing stable air-water linkage control. At the same time, the edge computing units deployed at each node of the tunnel communicate status information through a fixed-period broadcast mechanism, exchanging the local risk assessment results, the remaining resource status, and the node operation conditions every ten seconds. When any node experiences hardware failure or computational resource overload, the neighboring node starts the parameter migration mechanism after receiving its stop broadcast signal or status anomaly flag, and synchronizes and takes over the model parameters through the local cache and the central scheduling table. During the takeover process, the system selects the node with the lightest computational load as the takeover party according to the adjacency matrix and historical cooperation records to ensure that data processing and risk assessment are not interrupted, realizing dynamic self-adaptation of disaster recovery and fault tolerance.

[0050] Embodiment 2: This embodiment combines Figures 1 to 3 to illustrate the implementation of a smart fire safety warning system for highway tunnels. As Figure 1 shown, the system first receives video streams and sensor signals from the tunnel environment. These data are input into the data acquisition and preprocessing module. The data collected by the module includes video surveillance data, environmental sensor data, and vehicle positioning data, and the video pixel change rate and carbon monoxide concentration gradient are fused and analyzed to realize early smoldering feature recognition. On this basis, the system transmits the processed data to the dynamic risk assessment module to generate risk assessment parameters and the current gradient of the risk field model. This module constructs a fire risk propagation vector field using vehicle trajectories and smoke diffusion paths, and generates a dynamic heat map considering static parameters such as fire hydrants and smoke outlets and the dynamic distribution of vehicles. At the same time, the rescue window is deduced based on the game model of the fire spread speed and the traffic flow speed. The generated risk assessment results will be transmitted to the emergency resource elastic adaptation module, which automatically adjusts the induction, fire extinguishing, and smoke exhaust strategies according to the risk intensity, and outputs control instructions to the induction screen, pressure adjustment instructions to the fire pump, and speed control instructions to the smoke exhaust fan to realize intelligent regulation and linkage response of resources. At the same time, the edge computing nodes in the system cooperate with the data acquisition and preprocessing module to process the optical flow compensation data, ensure the stability and accuracy of the monitoring data, and take over tasks when a computing node fails to maintain the continuity of system operation. Finally, based on the fire extinguishing resource scheduling plan, the system can also control the path and spraying strategy of on-vehicle fire extinguishing devices to realize the coordinated operation of fixed facilities and mobile fire extinguishing equipment.

[0051] As Figure 2As shown in the figure, first, the monitoring terminal pushes environmental data (temperature / carbon monoxide / smoke) to the edge node; the edge node performs spatio-temporal data alignment to achieve precise temporal and spatial matching of video and sensor data; the edge node completes the transmission of risk parameters (the product of the pixel change rate and the concentration gradient > threshold 1), indicating that the early smoldering signs of a fire have been identified; subsequently, the decision-making center performs game strategy calculations based on the received risk parameters to construct a dynamic game model of the fire spread speed and the traffic flow speed; according to the calculation results, the decision-making center issues a control instruction set (pressure 80 to 150 MPa) to the actuator; to ensure execution accuracy, the edge node sends a status feedback request, thereby completing real-time working condition confirmation at the monitoring terminal; then, the actuator performs parameter fine-tuning based on the current feedback to ensure precise matching of pressure and smoke exhaust control with the actual situation; finally, the actuator returns an execution effect report (flow deviation less than 5%) to the decision-making center to verify the system response effect and form a closed-loop control.

[0052] As Figure 3 As shown in the figure, the edge computing network consists of Node 1, Node 2, Node 3, and Node 4. All nodes can interact with the data acquisition module and support collaborative operations with the dynamic risk assessment module; during the operation of the system, Node 1 synchronizes the model and risk field parameters to Node 3 through parameter migration, and Node 1 also transmits its local status and resource information to Node 2 in a periodic broadcast manner; when a single-point failure is detected in the system, Node 2 takes over the task to ensure that the core functions are not interrupted. At the same time, Node 3 and Node 4 share status information based on the collaborative perception mechanism to support model synchronization and fault judgment. During the task switching process, Node 4 can take over some computing tasks of Node 3 based on the seamless switching strategy to achieve dynamic migration and adaptive balance of the computing load during fault recovery. The entire edge computing framework ensures that the system can rely on node collaboration and task takeover mechanisms to maintain the steady state of the edge computing network in the event of emergencies or hardware anomalies, ensuring the continuous execution of dynamic risk assessment and the effective support of the data acquisition module.

[0053] Embodiment 3: For example, in a two-way six-lane highway tunnel located in mountainous terrain, to address the potential smoldering hazards that may occur during the peak period of night-time truck traffic, the system is deployed at a distributed edge node approximately 800 meters from the north entrance of the tunnel. This node is equipped with a high-definition video acquisition device, multiple groups of environmental parameter sensor arrays, a high-speed computing module, and a local induction control actuator. In the video surveillance section, the system uses video camera devices fixedly installed on the tunnel wall and facing the lanes to continuously collect image frame sequences and update the cache at a frequency of 25 frames per second. The image processing module uses a frame difference algorithm based on background modeling to extract the brightness change rate of each pixel on the time axis. In the calculation formula, the time window is selected as 2 seconds, and the pixel change rate is based on the average brightness increment between frames. The root mean square of the pixel gray-scale change is used as a local perturbation sensitivity index, and then clustering summation is performed on the image area to obtain the global pixel change rate estimate at the current moment. To avoid interference caused by single-point strong light or occlusion, the system divides the image into blocks and introduces a regional smoothing filtering mechanism. After removing the influence of abnormal noise, a stable time series of video pixel change rates is output. At the same time, the environmental parameter acquisition module continuously collects data from multiple carbon monoxide concentration sensors installed on the tunnel ceiling and both side walls, and organizes the data using a three-dimensional space grid structure. The system constructs a dynamic carbon monoxide concentration change gradient field through spatial interpolation and time fitting methods. The specific method is as follows: In the three-dimensional space grid constructed with sensors as nodes, each node uses a polynomial curve fitting within a sliding window in the time series to fit its concentration change trend. At the same time, bilinear interpolation is used in space to obtain the change rate in different directions, and combined into a set of local concentration change gradient vectors within a unit volume. The system synchronously samples the video pixel change rate and the carbon monoxide concentration change gradient field every 1 second and calculates their spatial consistency product. In the specific implementation, through a coordinate mapping mechanism, the corresponding monitoring area in the video image is paired with the nodes in the environmental parameter grid for point-to-point product calculation. When the product result continuously exceeds the empirically set response threshold (this threshold is based on the data cross characteristics in the first 15 seconds of 100 smoldering cases in the tunnel) for 5 seconds, the internal smoldering warning mechanism of the system is triggered.

[0054] In the coupled modeling of smoke propagation and vehicle trajectories, the system reconstructs the airflow velocity field based on the optical flow inversion mechanism. The specific operation is as follows: Select fixed light sources installed on the tunnel sidewall as reference points. By detecting the shape distortion and brightness distribution changes of these light sources in the video, the system uses the sparse optical flow method combined with template matching technology to extract optical flow vectors. After dimensionality reduction and spatial fitting of the optical flow field, it is mapped to the estimated airflow velocity values on the cross-section. Using this velocity field, the system non-linearly corrects the risk propagation path composed of vehicle trajectories and smoke boundaries, transforms the vehicle trajectories into reachable paths extending along the airflow direction, and establishes a vector layer superposition structure at multiple moments. The vehicle trajectory acquisition module is based on object detection and tracking technology in video images. It uses a lightweight convolutional neural network model to detect the vehicle body and realizes continuous tracking through inter-frame feature matching. The vehicle position point set is projected onto the corrected airflow coordinate system, and after superimposing the smoke boundary layer, a complete risk propagation vector field is formed. In each evaluation cycle, the system constructs a weighted reachability heat map based on the risk vector field, vehicle distribution map, and static layout of fire-fighting facilities (including the location and status of fire hydrants and smoke exhaust outlets). The generation method of the heat map includes: calculating the propagation path length from the current position of each vehicle to the potential fire source point, the average risk gradient of the area passed through, and the current fire resource reachability index, and using the combined weights of these factors as the basis for coloring the heat map. Further, to optimize the emergency resource dispatching, the system introduces a dynamic game model of traffic flow speed and fire spread speed. The traffic flow speed is obtained by continuously tracking the speed vector of the vehicle and averaging it in the lane area. The fire spread speed is comprehensively deduced from the smoke boundary expansion rate and local heat accumulation estimation. The two are used as the reachable window and the blockade diffusion interface respectively, constituting the limiting conditions for path feasibility in the model. The path optimization algorithm uses the variable-step difference method to deduce the rescue time window, and combines the vehicle steering limit conditions and the fire truck turning radius model to generate an optimal path set with continuous curvature. When generating the path set, the system considers the attenuation coefficient of the on-vehicle water tank capacity, assumes that the extinguishing agent decreases according to the distance linear loss model, and at the same time combines the position of the pipe network supply point to generate the fire extinguishing coverage area, thus forming a Pareto optimal path and pressure configuration solution set.

[0055] In the pollution perception and image restoration mechanism, the system automatically analyzes whether there is brightness abnormality or edge blur in the light source area of the monitoring image every 60 seconds. If it is detected that the brightness decrease amplitude exceeds 15% and the edge gradient is lower than the set clarity threshold within 5 consecutive cycles, the system automatically marks it as a polluted area. Combining the temporal consistency and the symmetry characteristics of adjacent light sources, the system extracts the corresponding area segments from the historical images of adjacent light sources and uses multi-frame fusion reconstruction technology to generate a restored image to replace the polluted area. This mechanism is immediately executed when it detects that the polluted area coincides with the vehicle prediction path, and the restoration result is fed back to the vehicle trajectory prediction module to update the prediction offset correction coefficient to avoid prediction distortion caused by image abnormalities.

[0056] In the linkage control mechanism between the water pump pressure and the smoke exhaust fan, the system collects the fire water pressure data in real time and calculates its second-order change amount within a 2-second sliding window to determine whether there is a tendency of pressure mutation. When the second-order change value exceeds the threshold (this value is obtained through tunnel fluid model simulation and has been preset as different time limits in different cross-section structures), the system immediately adjusts the speed of the smoke exhaust fan through the proportional control mechanism to suppress the backflow turbulence effect caused by the high-pressure water column. In the control logic, the second-order change value is used as the feedback input of the fan speed governor, and the current air humidity and temperature indicators are combined to form a comprehensive response factor to ensure a balance between the sensitivity and stability of the control strategy. At the same time, to ensure the continuous operation and fault tolerance of the system, a periodic status broadcast mechanism is established between the edge computing modules at each node in the tunnel. The local risk assessment results, device status, and computing load information are synchronized every 10 seconds. When a certain node fails to broadcast periodically or sends an abnormal status flag, the adjacent node immediately starts the model parameter migration mechanism based on the topology table and load priority, and seamlessly transfers the current computing task to the node with the minimum load and a historical stable record to achieve function takeover. The takeover process includes steps such as parameter status reconstruction, interrupted frame recalculation, and interface address mapping replacement, etc., to ensure that the overall system evaluation is not interrupted and the response strategy is not aborted during the task conversion, improving the robustness and reliability of the system in the complex environment of long tunnels. All of these belong to the extended implementation methods known to those of ordinary skill in the art.

[0057] Example 4: In terms of constructing the risk propagation vector field, the system uses a video frame sequence of five consecutive seconds as the input window, and uses a pixel trajectory extraction algorithm based on foreground motion area segmentation to identify the shape change trend of the vehicle edge and the spatial expansion path of the smoke texture in each frame. The extracted edge information forms a texture gradient map after non-maximum suppression, and then candidate boundary lines are generated by the local maximum tracking algorithm. The system establishes an inter-frame mapping relationship between the boundary lines and the recognized vehicle contours through the feature point tracking algorithm. On this basis, a Kalman filter is used to construct the vehicle prediction trajectory curve, and a reference coordinate system is defined with this curve as the main axis. The direction of each smoke boundary expansion vector is determined by the boundary difference vector between the current frame and the previous frame, and the length is weighted according to the local texture gray change rate. The finally generated vector field covers the entire effective area of the image. At the same time, to correct the offset error of the above vector field under the airflow disturbance, the system estimates the air flow velocity on the current cross-section in real time based on the optical flow inversion mechanism of the fixed light source area. In specific implementation, the system selects two groups of fixed light sources with equal distances on the tunnel wall, and extracts the brightness change and edge displacement information in consecutive frames through template matching and sparse optical flow field estimation method. The system uses Newton interpolation method to construct the brightness change surface, and converts the gradient component in the horizontal direction into the air flow velocity estimation value. This velocity value is extended spatially through bilinear interpolation to form a two-dimensional air flow velocity field on the cross-section. This velocity field then acts on the smoke expansion vector field, the vector direction is rotationally compensated according to the local airflow dominant direction, and the length is scaled according to the airflow velocity distribution to form a corrected risk propagation vector layer.

[0058] The system superimposes the vector layer and the vehicle prediction trajectory to construct a risk propagation path set, and calculates the area range covered by the risk propagation along each vehicle trajectory direction within a unit time step starting from the initial fire point. Each vehicle path point will be assigned a propagation weight value according to the vector density, superposition direction consistency and historical risk record at its location to reflect the intensity distribution of the risk field. The risk values of all path points are processed by Gaussian blur to form a risk heat map, providing a quantitative basis for subsequent emergency resource scheduling. In the pollution image recognition and repair mechanism, the system performs a brightness histogram analysis on the light source area in the image every ten seconds, extracts the brightness peak, edge contrast and clarity indicators. If any indicator drops by more than 15% within three consecutive cycles, the system determines that this area may be polluted. At this time, the image repair module is started, and the symmetric area segment is extracted from the historical images of adjacent light source areas, and pixel-level repair is performed using the weighted fusion algorithm combined with the texture information of the pollution-free area in the current frame. After the repaired image undergoes edge reconstruction and gray level equalization processing, it is transmitted back to the vehicle trajectory prediction module to update its background modeling parameters to prevent the accumulation of trajectory errors caused by pollution.

[0059] In terms of the game modeling of fire spread and traffic flow, the system takes the current position of the vehicle, the predicted speed, and the current risk path map as inputs. In each evaluation cycle, it constructs the expansion boundary of the fire source influence area. The fire spread speed is jointly derived from the average displacement speed of the smoke boundary in consecutive frames and the local temperature sensor increment curve. The traffic flow speed is calculated from the average instantaneous speed of the vehicle ahead. The two respectively form the risk domain boundary and the vehicle passable area. Based on this, a path reachability matrix is constructed. Each candidate path will consider constraints such as the vehicle turning radius, passing width, and fire extinguishing resource delivery ability. The multi-objective differential evolution algorithm is used to solve the path set with the shortest response time and the minimum attenuation of the aqueous agent. The aqueous agent attenuation factor is estimated based on the linear function between the water volume carried by the vehicle and the path length, and water supply replenishment nodes are generated in combination with the fixed fire hydrant distribution to make the path optimization result physically realizable. Finally, the edge computing nodes achieve model synchronization and task takeover functions through status broadcasts with a 5-second cycle. Each node maintains the local topology structure diagram and the cache of adjacent node parameters. When it detects that an adjacent node stops broadcasting or sends an abnormal status flag, the system selects a takeover node according to the historical calculation stability and the current calculation load priority, and completes the seamless migration of the calculation task through three stages: parameter synchronization, status reconstruction, and task switching. The parameter synchronization stage includes the update of the risk layer status, the heat map distribution, and the vehicle prediction parameters. The status reconstruction stage performs the backfill of the unfinished data. The task switching stage ensures business continuity through interface address mapping. The whole mechanism ensures that when any node fails, the overall operation logic of the system does not interrupt, and the continuity of risk assessment and control response is guaranteed. These are all extended implementation methods known to those of ordinary skill in the art.

[0060] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent fire safety warning system for highway tunnels, characterized in that, The system includes: A data acquisition and preprocessing module is used to collect multi-source heterogeneous data in the tunnel in real time. The multi-source heterogeneous data at least includes video surveillance data, environmental sensor data, and vehicle positioning data. It extracts vehicle movement trajectories and smoke texture features from the video surveillance data, performs spatio-temporal interpolation compensation on the environmental sensor data, and calculates the cross-modal early smoldering feature recognition for the video pixel change rate ΔP in the video surveillance data and the carbon monoxide concentration change gradient in the environmental sensor data. When the following conditions are met it is determined as early smoldering. A dynamic risk assessment module, which is communicatively connected to the data acquisition and preprocessing module, and is used to construct a fire risk field model based on the spatio-temporal coupling of multi-source heterogeneous data. Specifically, it includes: generating a risk propagation vector field based on vehicle movement trajectories and smoke texture features, combining static parameters such as the positions of fire hydrants and the states of smoke exhaust outlets in the tunnel and the real-time vehicle distribution to generate a dynamic reachability heat map, and establishing a game model between the fire spread speed and the traffic flow speed to predict the optimal rescue time window; An emergency resource elastic adaptation module, which is communicatively connected to the dynamic risk assessment module, and is used to automatically switch the display strategy of the induction screen according to the intensity gradient of the fire risk field model, and form a negative feedback closed loop for the pressure regulation of the fire pump and the rotation speed of the smoke exhaust fan; Edge computing nodes, which are distributed in the tunnel, are used to execute the computing tasks of the data acquisition and preprocessing module and the dynamic risk assessment module, and realize the adaptive transfer learning of model parameters through topology awareness, and seamlessly take over the computing tasks of adjacent nodes in case of single-point failure.

2. The intelligent fire safety early warning system for highway tunnels according to claim 1, wherein The data acquisition and preprocessing module also includes an optical flow compensation mechanism, specifically: using the optical flow field of a fixed light source in the tunnel sidewall monitoring video to invert the real-time air flow velocity field, and dynamically correcting the spatio-temporal mapping relationship between the smoke diffusion path and the vehicle trajectory based on the real-time air flow velocity field to compensate for the non-linear spatio-temporal distortion caused by air flow disturbance.

3. The intelligent fire safety early warning system for highway tunnels according to claim 2, characterized in that, The data acquisition and preprocessing module also includes a local pollution perception and image self-repair mechanism, specifically including: identifying the pollution characteristics on the surface of the fixed light source based on a lightweight model to generate a pollution area mask; constructing a multi-frame reference repair matrix for the pollution area by using the spatial symmetry of adjacent fixed light sources and the temporal illumination consistency; and when the ratio of the pollution area S polluted / S total > θ2 and lasts for a preset time T polluted , activating the repair algorithm and suppressing non-emergency repairs according to the humidity data in the tunnel, where S polluted is the area of the pollution area, and S total is the total area of the light source.

4. The intelligent fire safety warning system for highway tunnels according to claim 3, characterized in that, The local pollution perception and image self-repair mechanism also includes: when it is recognized that the pollution area overlaps with the predicted path of the vehicle trajectory, giving priority to repairing this area; dynamically adjusting the sampling frequency of the repair area based on the predicted trajectory speed; and the repaired optical flow data is fed back to optimize the vehicle trajectory prediction model.

5. The intelligent fire safety early warning system for highway tunnels according to claim 1, wherein, When the emergency resource elastic adaptation module forms a negative feedback closed loop for the pressure regulation of the fire pump and the rotation speed of the smoke exhaust fan, a second-order derivative feedback term of the fire water pressure sensor is embedded in the control loop of the smoke exhaust fan to dynamically suppress the air turbulence caused by sudden changes in water pressure.

6. The intelligent fire safety warning system for highway tunnels according to claim 1, characterized in that, The edge computing nodes broadcast their local risk assessment results and resource status information periodically to achieve collaborative perception of the overall risk situation of the tunnel and distributed collaborative scheduling of emergency resources.

7. The intelligent fire safety early warning system for highway tunnels according to claim 1, wherein, When the dynamic risk assessment module generates the risk propagation vector field, it aligns the vehicle movement trajectory in the video stream with the smoke diffusion direction at the microsecond level to compensate for the temporal misalignment between the smoke diffusion path and the video movement trajectory caused by tunnel air flow disturbance.

Images