Power plant fire intelligent linkage disposal system

Abstract

The invention discloses a power plant fire intelligent linkage disposal system, which comprises a high-precision sensing layer, a time sequence prediction engine, a game decision system and an execution control system, and is characterized in that the high-precision sensing layer collects original data through a distributed sensor, and the original data is preprocessed into 12-dimensional feature vectors through edge computing nodes; the time sequence prediction engine takes the 12-dimensional feature vector as input, adopts a physical constraint Transform model, embeds a heat conduction equation regularization item, and outputs a fire evolution probability graph; the game decision-making system generates an optimal evacuation path through a multi-target reinforcement learning algorithm based on a fire evolution probability graph and a feature vector, and generates an equipment control instruction through a three-level response mechanism; and the execution control system drives an executor to execute a control instruction through a time-sensitive network, and feeds back an equipment state to the game decision-making system in real time. According to the invention, all links are mutually linked and dynamically adjusted, and efficient cooperation and accurate response of fire monitoring, early warning, decision making and disposal are realized.

Description

Technical Field

[0001] The present invention relates to the field of industrial safety technology, and in particular to an intelligent linkage handling system for power plant fires. Background Art

[0002] Existing fire prevention and control technologies in industrial sites like power plants present numerous challenges, significantly impacting their effectiveness and safety. Traditional fire monitoring systems typically utilize a single threshold alarm mechanism, triggering an alarm when the set temperature exceeds 70°C, for example. However, power plants operate under extremely complex conditions, with frequent non-fire conditions such as heat dissipation in steam pipes and abnormal heating of equipment. This leads to frequent false alarms in these monitoring systems, which not only consumes significant manpower and resources to investigate and resolve, but also can lead to staff becoming indifferent to the alarms.

[0003] Currently, static plans are often used for evacuation route planning. However, fires spread dynamically, making changes in fire intensity difficult to predict. Pre-set evacuation routes are often unavailable due to sudden changes in the fire's intensity, posing a significant threat to personnel safety. Once a fire breaks out, people are likely to be put in danger due to untimely evacuation.

[0004] Interoperability between fire suppression equipment, smoke exhaust systems, and access control is also a major challenge. NIST testing shows that the average delay between the issuance of an alarm signal and the activation of relevant equipment in existing systems exceeds 150 seconds. In a fire emergency, this 150-second delay can cause the fire to spread rapidly, missing the optimal time to extinguish the fire and causing greater losses.

[0005] Faced with these problems, researchers have attempted various improvements. For example, multi-sensor fusion technology was used to improve the accuracy of fire identification. However, because it lacked a physical environmental model, predictions of fire spread were significantly biased, making it difficult to accurately determine fire trends. Alternatively, reinforcement learning was used to optimize path planning. However, this lacked a dynamic obstacle avoidance mechanism, making path conflicts prone to occur in practical applications, impacting evacuation efficiency and safety.

[0006] Furthermore, existing technologies commonly suffer from several key unresolved pain points. Most algorithms rely solely on data-driven approaches, failing to incorporate physical laws such as thermodynamic equations. This results in fire predictions that differ significantly from actual fire development. Furthermore, the separation of path planning and equipment control results in low efficiency during multi-objective optimization, hindering rapid and effective decision-making. The lack of a unified scheduling protocol between actuators, such as firefighting robots and sprinkler systems, makes response delays prone to resource competition during task execution, hindering the formation of an efficient collaborative working mechanism. Summary of the Invention

[0007] In order to solve the problems raised in the above background technology, the present invention provides an intelligent linkage disposal system for power plant fires, which aims to solve the core problems existing in the existing power plant fire prevention and control: to address the frequent false alarm problem caused by the single threshold alarm of the traditional monitoring system, through the multi-source data fusion and physical constraint Transformer model of the high-precision perception layer, physical laws such as thermodynamic equations are embedded to improve the ability to distinguish between fire and non-fire conditions and reduce the false alarm rate; to address the defect that static evacuation path planning cannot adapt to the dynamic changes of the fire, a dynamic path planning module is used to combine the real-time fire evolution probability map and the dynamic characteristics of personnel to generate three optimal paths every 15 seconds, dynamically avoid risk areas such as high temperature and low visibility, and improve evacuation efficiency and safety; to address the problem of high equipment coordination delay, a three-level response mechanism is constructed through time-sensitive network and preemptive scheduling protocol to achieve low delay from alarm to equipment linkage, and dynamically allocate resources through risk weights to resolve competition conflicts between actuators such as fire-fighting robots and sprinkler systems, thereby improving collaborative disposal efficiency. The system realizes efficient coordination of fire monitoring, early warning, decision-making and disposal through the closed-loop linkage of the entire process of "perception-prediction-decision-making-execution", effectively curbing the spread of fire and reducing casualties and property losses.

[0008] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0009] An intelligent linkage fire handling system for power plants includes a high-precision perception layer, a time series prediction engine, a game decision system, and an execution control system. The high-precision perception layer consists of a distributed infrared temperature sensor array, a multi-plane gas monitoring unit, and an optical fiber deformation sensor. The raw data collected by each sensor is preprocessed by an edge computing node to generate a 12-dimensional feature vector consisting of temperature, gas concentration, deformation rate, and personnel dynamic characteristics within a continuous time window, providing standardized input for the time series prediction engine. The time series prediction engine adopts a physical constraint Transformer model, takes the 12-dimensional feature vector as input, and outputs a spatially gridded fire evolution probability map, which is synchronously transmitted to the game decision system. The game decision system includes a dynamic The path planning module and the hierarchical equipment linkage controller, the dynamic path planning module uses a 15-second cycle, combines the environmental parameters and personnel dynamic characteristics in the fire evolution probability map, and generates three optimal evacuation paths through a multi-objective reinforcement learning algorithm. The hierarchical equipment linkage controller has a built-in three-level response mechanism, which generates equipment control instructions based on the fire evolution probability map and the 12-dimensional feature vector, and performs hierarchical control of the equipment; the execution control system receives the optimal evacuation path and equipment control instructions through a time-sensitive network to drive the actuators including the fire robot cluster, variable sprinkler system, and intelligent smoke exhaust device to execute the equipment control instructions, adopts a preemptive scheduling protocol to coordinate the actions of multiple devices, and feeds back the equipment status to the hierarchical equipment linkage controller in real time for dynamic adjustment of PID parameters or path switching.

[0010] Preferably, the raw data of the infrared temperature sensor array is a discrete temperature value collected by the infrared temperature sensor. The infrared temperature sensor array is deployed according to a hexagonal grid topology, and the distance d between adjacent nodes is between 6m and 8m. Such a layout can ensure efficient perception of the thermal field in key areas including cable mezzanines and steam turbine rooms.

[0011] Preferably, the sensors used in the multi-plane gas monitoring unit include a CO sensor and The sensor is set up with three detection layers in the vertical direction. The installation height h is proportionally distributed based on the site height H. The height of the bottom detection layer is , the height of the middle detection layer , top detection layer height ,The raw data collected by the multi-plane gas monitoring unit are the ,gas concentration values ​​collected by the sensor at a sampling frequency of 10 Hz.

[0012] Preferably, the optical fiber deformation sensor adopts an FBG array structure, and the original data is the wavelength drift based on the FBG array. , each sensing unit contains 4 Bragg gratings.

[0013] Preferably, the data preprocessing uses radial basis function neural network to perform spatial interpolation on discrete temperature values ​​collected by infrared temperature sensor array to construct precision Continuous temperature field model Generate continuous temperature field values, calculate spatial partial derivatives based on the continuous temperature field values ​​using Fourier's law, and derive the temperature gradient vector , and combined with the material thermal conductivity k, through the formula Calculate the heat flux density q to characterize the direction and intensity of heat conduction. At the same time, ensure the time synchronization of temperature data through the time synchronization protocol, and derive time series characteristics including temperature change rate. ; For the gas concentration values ​​collected by the multi-plane gas monitoring unit, the filtering technology is used to suppress the airflow disturbance noise and output the denoised CO, Concentration value, based on the gas concentration values ​​of the three detection layers, the gas diffusion gradient is calculated using Fick's law, and the gas concentration values ​​of the three detection layers are normalized to a unified coordinate system to generate vertical concentration distribution characteristics; for the wavelength drift collected by the optical fiber deformation sensor Then it is converted into structural deformation through the Bragg grating formula ,in is the refractive index of light, The grating period can monitor the deformation of the load-bearing beam in real time, and then calculate the deformation rate within the time window through the sliding window algorithm. Evaluate the dynamic trend of the structure, and finally perform threshold comparison to calculate the deviation between the structural deformation and the safety threshold to generate risk warning characteristics; at the same time, multi-sensor data fusion technology is used to analyze the dynamic characteristics of personnel, based on the continuous temperature field model In the gradient distribution of the spatial dimension and the dynamic evolution of the time dimension, the moving trajectory and speed of the high-temperature area are calculated by the motion target detection algorithm to extract the moving speed of the hot zone reflecting the movement of personnel. Then, the abnormal fluctuation of gas concentration value is analyzed by CO concentration value. When the CO concentration shows a sudden change of non-natural diffusion, the degree of disturbance of air flow by crowd flow is inferred based on the fluid mechanics model, and the crowd flow index is calculated. Finally, the parameters are fused to generate the hot zone moving speed and crowd mobility index The dynamic characteristics of personnel are input into the time series prediction engine to predict behavior patterns.

[0014] Preferably, the 12-dimensional feature vector includes temperature, gas concentration, deformation rate and personnel dynamic characteristics, and the temperature feature includes continuous temperature field value, temperature gradient vector , heat flux density q, time series characteristics, the gas concentration characteristics include three detection layers of CO and Concentration value, gas diffusion gradient, vertical concentration distribution characteristics, the deformation rate characteristics include structural deformation , deformation rate , risk warning characteristics, the personnel dynamic characteristics include the speed of movement in the hot zone and population mobility index ,The above temperature, gas concentration, deformation rate and personnel dynamic ,characteristics are normalized to the [-1,1] interval through the z-score, ,forming a 12-dimensional feature vector that meets the input requirements of the ,time series prediction engine.

[0015] Preferably, the physical constraint Transformer model comprises a 6-layer stacked structure, each layer is configured with 4 independent attention heads, which can effectively extract multi-dimensional data features, wherein the 2nd and 4th layers introduce physical constraint modules, and the heat conduction equation is used to calculate the attention head. A regularization term is constructed, where α is the thermal diffusivity of the material and Q is the heat source term, so that the prediction of the physical constraint Transformer model is more consistent with physical laws and significantly improves the accuracy of the prediction.

[0016] Preferably, the pre-training stage of the physical constraint Transformer model uses a power plant fire dataset containing 500,000 groups of samples. During transfer learning, the encoder parameters of the first three layers are frozen, and only the last three layers and the physical constraint module are fine-tuned to improve training efficiency while avoiding overfitting.

[0017] Preferably, the loss function of the physical constraint Transformer model is defined as , β=0.3,γ=0.1 are determined by cross-validation optimization, where is the L2 regularization term, which further improves the generalization ability of the physical constraint Transformer model.

[0018] Preferably, the physical constraint Transformer model also adopts a sliding window online learning mechanism to update the model weights every 15 minutes. When it is detected that the operating condition changes exceed the threshold, the model update is triggered in advance to ensure that the model can adapt to the changing operating conditions of the power plant in a timely manner, and finally output a spatial gridded fire evolution probability map to provide an accurate prediction basis for subsequent decision-making.

[0019] Preferably, in order to ensure the reliability of the output of the physical constraint Transformer model, the time series prediction engine has built an expert rule library containing more than 200 thermodynamic criteria to verify the model output. Once the model prediction conflicts with the physical rules, the double verification mechanism is immediately activated to ensure the credibility of the prediction results.

[0020] Preferably, the dual verification mechanism is that when the fire evolution probability map output by the physical constraint Transformer model conflicts with more than 200 thermodynamic criteria in the expert rule base, including heat conduction equations, gas diffusion laws and other physical rules, the system first calls another set of physical models independent of the original physical constraint Transformer model, inputs the same 12-dimensional feature vector for recalculation, and generates comparative fire prediction results. Secondly, the manual intervention interface is triggered to synchronously push the model prediction results and the dual physical model calculation results to the power plant safety monitoring center, and the expert team conducts multi-source data cross-validation. By analyzing the consistency of the spatiotemporal distribution of temperature fields, gas concentration fields, and deformation rates, the credibility of the prediction results is finally determined to ensure that the prediction results are in line with actual physical laws under complex working conditions and avoid decision-making risks caused by misjudgment of a single model.

[0021] Preferably, the environmental parameters obtained by the dynamic path planning module based on the fire evolution probability map include the temperature gradient vector ,visibility , Probability of passing through the escape route Construct a three-dimensional thermal risk assessment system based on temperature gradient vector Defining the temperature risk field ,when Corresponding gradient value When , the path planning algorithm automatically avoids the area. If it must pass through, the angle between the path and the heat flow direction is forced to be , and mark "Quick Pass" prompts at the corners of the path; visibility The real-time calculation is based on the CO concentration value of the multi-plane gas monitoring unit, and the mapping relationship between CO concentration and visibility is established through the Beer-Lambert law. ,in The standard visibility for smokeless fire scene is 50m. is the attenuation coefficient of CO to visible light , C is the CO concentration measured by the multi-plane gas monitoring unit, L is the length of the planned path, when visibility When the distance is less than 3 meters, the path is prioritized to be close to the wall or in the area with dense evacuation lights, reducing the risk of people getting lost through visual guidance; the probability of passing through the escape route The assessment is based on the structural safety level S of the optical fiber deformation sensor, the status of the fire protection facilities, including the opening and closing status of the fire door , Fire hydrant pressure , and the number of times N that the channel is blocked in the past 5 minutes are used to perform Bayesian network modeling. The probability model is ,in is the Sigmoid function, mapping the linear combination to interval, dynamically predict channel blocking probability, when When the channel is marked as "high risk congestion", the weight in path planning is reduced by 60%.

[0022] Preferably, the structural safety level S of the optical fiber deformation sensor is determined according to the structural deformation amount. and deformation rate The dual-indicator judgment divides the structural safety status into 5 levels. The edge computing node updates the structural safety level S every 10 seconds and synchronizes it to the dynamic path planning module and the hierarchical equipment linkage controller through the time-sensitive network.

[0023] Preferably, when When S=1; when When S=2; when When S=3; when When S=4; when When S=5.

[0024] Preferably, the number of times N that the channel is blocked in the past 5 minutes adopts a sliding time window with a duration of 5 minutes and a sliding step of 30 seconds, and the number of times N that the blocking events occur in the channel is counted in real time. The determination of the blocking event is carried out by combining the data fusion of the infrared temperature sensor array and the multi-plane gas monitoring unit. When the density of people in a certain channel area is , and the CO concentration gradient fluctuation amplitude , it is determined to be a "congestion event", in which the density of people By hot zone area and temperature threshold Cross-validation obtained.

[0025] Preferably, the hot zone area is the temperature The area of ​​the area; the population density is the number of people per unit area, which is calculated by the ratio of the hot zone area to the preset space area; the CO concentration gradient fluctuation amplitude is the CO concentration gradient The difference between the maximum and minimum values ​​within the time window reflects the intensity of the airflow disturbance.

[0026] Preferably, the dynamic path planning module constructs a dynamic group characteristic model containing spatiotemporal distribution parameters based on the dynamic characteristics of personnel to perform dynamic group behavior analysis, wherein the hot zone moving speed The moving target detection algorithm is used to track the centroid of the moving hot spot in the continuous temperature field model, and the displacement of the hot spot centroid per unit time is calculated to quantify the personnel movement speed characteristics. When the fast-moving crowd is judged to include young and middle-aged people, the shortest time path is assigned first; when the hot zone moving speed is When the slow-moving crowd is determined to include the elderly, children and people with mobility difficulties, the path planning will give priority to matching safe routes with low slopes and wide passages, and the fire-fighting robot will be linked to guide and escort them.

[0027] Preferably, the crowd mobility index The calculation first calculates the CO concentration gradient Perform wavelet decomposition to extract high-frequency components with frequencies > 2Hz and low-frequency background components , and then through the perturbation intensity formula Quantifying the impact of population mobility When the crowd is divided into Divided into three teams: fast, medium and slow, they are directed to independent paths with a spacing of ≥5m, including the fast team taking the main passage and the slow team taking the spare stairs; When a fire occurs, the "stagnation warning" response is activated, and the only goal of path planning is to reach the nearest exit within 5 minutes. At the same time, a fire robot cluster instruction is sent to the execution control system, and the fire robot leads the way to open up channels and provide lateral protection to ensure safety.

[0028] Preferably, the multi-objective reinforcement learning algorithm integrates a three-dimensional thermal risk assessment system, fits the strategy function through a 6-layer fully connected network, constructs a Markov decision process model, and generates three optimal paths every 15 seconds, including the shortest time path, the lowest risk path, and the high fault tolerance path, among which the shortest time path focuses on For fast-moving people, the path length optimization weight accounts for 60%, allowing crossing The medium-risk hotspot; the lowest-risk path targets Slow moving crowds, avoiding all Area, give priority to passing through equipment supply points with a distance of ≤15m; high fault tolerance path is suitable for In congested scenarios, the path contains ≥2 backup channels, allowing a maximum detour distance of 1.5 times the shortest path to ensure evacuation redundancy under extreme conditions.

[0029] Preferably, the multi-objective reinforcement learning algorithm uses the Markov decision process to model the dynamic fire scene, abstracts the environmental parameters and personnel dynamic characteristics into the state space of MDP, defines operations such as path selection and direction adjustment as the action space, and constructs a dynamic mapping relationship of "state-action-reward".

[0030] Preferably, based on this modeling, the multi-objective reinforcement learning algorithm performs the following steps in each cycle: first, the parameters including the real-time temperature field, gas concentration field, structural safety level, and crowd distribution are encoded as a state vector, and then based on the preset reward function, the 6-layer fully connected network parameters of the strategy function are optimized by the temporal difference algorithm to maximize the cumulative reward, and finally output three optimal paths adapted to different scenarios.

[0031] Specifically, the reward function is the shortest time path reward is negatively correlated with the path length, and the penalty is negatively correlated with the path length. High temperature area contact time, minimum risk path reward and avoidance The degree of high-risk area is positively correlated, the penalty is related to the amount of exposure to toxic gas, the reward for high-tolerance paths is positively correlated to the number of backup channels, and the penalty is related to the proportion of the detour distance exceeding the shortest path.

[0032] Preferably, the control logic for triggering the three-level response mechanism and generating device control instructions is as follows.

[0033] Preferably, in the first level response, the hierarchical equipment linkage controller analyzes the fire evolution probability map output by the time series prediction engine in real time to extract the fire probability ,when And duration When the edge computing node is used, the following instructions are executed: control instructions are sent to the power monitoring system in the actuator through the edge computing node to shut down the power supply of non-critical areas, including the non-emergency circuit of the lighting system, to reduce the risk of electrical short circuit; the sound and light alarm system in the actuator is driven to issue intermittent alarms, and voice guidance is played through the fire alarm broadcast; the hot zone movement speed based on the 12-dimensional feature vector is detected. ,like The evacuation indicator light is controlled by the execution control system to indicate the temporary evacuation route by changing the flashing frequency.

[0034] Preferably, the fire probability The probability value of the fire evolution probability diagram in the next 1 minute is the probability value of the fire. The value range is , calculated by the physical constraint Transformer model based on the 12-dimensional feature vector; the duration is the duration that the fire probability exceeds the threshold continuously.

[0035] Preferably, the trigger condition for the secondary response is the temperature change rate in the 12-dimensional feature vector or CO concentration , and the diffusion rate of high temperature area in the fire evolution probability diagram At this time, the grading equipment linkage controller executes the following instructions: send instructions to the intelligent smoke prevention and exhaust device, adjust the fan operation state to the medium speed gear, and give priority to eliminating the middle detection layer The fire door linkage module of the game decision system is controlled through the time-sensitive network to automatically lower the fire doors of the three adjacent fire partitions in the fire area, and mark the isolation area on the map module of the game decision system; according to the multi-plane gas monitoring unit Gas concentration value, if A cut-off command is sent to the fuel supply system of the execution control system to close the electromagnetic valve to block the hydrogen delivery.

[0036] Preferably, the diffusion rate of the high-temperature region is calculated by the spatiotemporal gradient of the continuous temperature field model, and is defined as the displacement velocity of the center of mass of the region with a temperature ≥ 80°C per unit time, and the calculation formula is: ,in, is the coordinate of the centroid of the high temperature area at time t, It is the output period of the time series prediction engine. This parameter is used to characterize the dynamic trend of fire spread.

[0037] Preferably, the triggering of the three-level response must simultaneously meet the highest temperature in the fire evolution probability map. , deformation rate And the population mobility index , the hierarchical equipment linkage controller sends the highest priority equipment control instructions through the time-sensitive network, including: sending partition control instructions to the variable sprinkler system, according to the temperature gradient Dynamically adjust the nozzle flow, including The regional nozzle pressure is increased to 1.2 times the rated value, forming a continuous water curtain isolation zone; the fire-fighting robot cluster is dispatched, guided by the high-tolerance path generated by the dynamic path planning module, combined with the structural safety level S of the fiber optic deformation sensor to avoid high-risk areas, and the fire extinguishing agent is sprayed after reaching the fire source; the emergency lighting system in the actuator is activated to force switching to the backup power supply to ensure that the illumination of the evacuation channel is ≥10lx, and the coordinates of the non-evacuated personnel are marked in real time based on multi-sensor data fusion technology, and coordinated rescue instructions are sent to the nearest fire-fighting robot.

[0038] Preferably, the hierarchical equipment linkage controller dynamically adjusts PID parameters or path switching based on equipment status deviation by executing the equipment status fed back in real time by the control system. The specific implementation method is as follows.

[0039] Preferably, when the control system is executed to feedback the pipe water pressure value of the variable sprinkler system and target value Deviation exceeds When the controller extracts the deviation signal and rate of change , automatically adjust the proportional coefficient through PID algorithm :like and , when the water pressure continues to be lower than the target value, Increase the pressure in steps of 0.1 to 1.2 to speed up the boost response; if and , when the water pressure is close to stable, It decreases to 0.6 to suppress overshoot; at the same time, when the water pressure deviation continues to fail to converge for more than 30 seconds, the standby pump start command is automatically triggered to ensure fire extinguishing efficiency.

[0040] Preferably, for the real-time position feedback of the fire-fighting robot cluster, if it is detected that a robot deviates from the optimal path generated by the dynamic path planning module by more than 0.5m, and the optical fiber deformation sensor does not feedback the structural deformation warning of the area, the hierarchical equipment linkage controller immediately selects the suboptimal path with the shortest Euclidean distance to the current position from the three pre-generated optimal paths, sends a path replanning instruction to the robot, and adjusts the PID parameters of its drive motor at the same time: increase the differential coefficient To enhance path tracking rigidity, ensure return to the new path within 15 seconds, and path switching delay ≤ 1.5 seconds.

[0041] Preferably, when the intelligent smoke exhaust device feedback fan speed fluctuation exceeds the rated value When the grading equipment linkage controller is combined with the CO concentration gradient of the multi-plane gas monitoring unit Make a joint judgment: If the CO concentration gradient If the CO concentration gradient exceeds 0.3ppm / m, it indicates abnormal airflow disturbance and is judged to be caused by pipeline blockage, resulting in abnormal speed. A shutdown command is immediately sent to trigger the start of the backup fan, and a mechanical clearance request is sent to the execution control system. Normal, recalculate the integral coefficient through the PID parameter self-tuning algorithm , suppress the speed fluctuation to range to ensure stable smoke exhaust efficiency.

[0042] Preferably, during the process of dynamically adjusting the PID parameters, the hierarchical device linkage controller records the feedback device status in real time and uploads it to the timing prediction engine through the edge computing node as a training sample for the physical constraint Transformer model to optimize the prediction accuracy of future control instructions.

[0043] Preferably, the preemptive scheduling protocol constructs a dynamic priority allocation mechanism based on risk level and task urgency to achieve high efficiency and reliability of actuator collaborative control.

[0044] Preferably, the risk level is divided into three levels: low, medium and high based on typical fire scenarios in power plants. The risk level is evaluated by combining the digital twin model with parameter thresholds including temperature, gas concentration and structural deformation, and the weight coefficient is optimized through 2000 simulation tests. , particle swarm optimization algorithm is used to iteratively optimize the weight coefficient , and finally determine the first-level priority corresponding to the high-risk scenario , Level 2 priority corresponds to medium-risk scenarios , Level 3 priority corresponds to low-risk scenarios .

[0045] Preferably, the dynamic priority allocation mechanism includes a three-level priority classification system: the first level priority, the fire extinguishing instruction of the fire robot cluster, is used for open fire fighting or structural safety level In high-risk scenarios, it can interrupt other non-critical tasks and monopolize communication bandwidth to ensure rapid arrival at the fire source; secondary priority, partition control instructions for variable sprinkler systems, when the temperature gradient vector or When the fire is in the vicinity of the fire source, its priority is automatically raised to the first level to ensure the rapid start of the water curtain isolation around the fire source; the third level priority, the ventilation adjustment instruction of the intelligent smoke exhaust device has the lowest default priority, but in the case of visibility And the population mobility index When the system is in operation, based on the multi-sensor data fusion results including the heat zone density of the infrared temperature sensor array, its priority is automatically raised to level two to ensure air circulation in the evacuation channel.

[0046] Preferably, the dynamic priority allocation mechanism also includes a dynamic preemption mechanism, which implements real-time scheduling through edge computing nodes and calculates the real-time risk weight of each executor every 5 seconds. , when the real-time risk weight of an actuator Value exceeds threshold Emergency preemption is triggered.

[0047] Preferably, to address the resource competition problem among multiple executors, the preemptive scheduling protocol adopts a risk and timeliness dual-factor arbitration strategy. When multiple executors request the same resource, the arbitration score formula is used: Conduct an assessment, where is the current risk value of the device, For the remaining time of the task, the person with the highest score will gain control of resources, ensuring that high-risk, short-term tasks are executed first, effectively avoiding response delays caused by resource competition, and improving the efficiency of collaborative processing of multiple devices.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. This invention uses a physically constrained Transformer model to embed physical laws such as the heat conduction equation as regularization terms into the prediction process, effectively filtering out interference from non-fire heat sources. Through an expert rule base and a dual verification mechanism, it effectively distinguishes between fire and non-fire conditions, significantly reducing the probability of false alarms.

[0050] 2. Based on the fire evolution probability map output by the physical constraint Transformer model and integrating the dynamic characteristics of personnel, this invention generates three optimal evacuation routes every 15 seconds through a multi-objective reinforcement learning algorithm. Simultaneously, a three-dimensional thermal risk assessment system analyzes environmental parameters in real time, dynamically avoids high-risk areas, and provides differentiated route planning for different groups of people, improving evacuation safety and efficiency and reducing route conflicts.

[0051] 3. This invention builds a three-level response mechanism through time-sensitive networking and preemptive scheduling protocols. It combines a dynamic priority allocation mechanism with an arbitration scoring formula to optimize resource scheduling, shorten device linkage delays, improve collaborative response efficiency, and effectively curb the spread of fire.

[0052] 4. This invention uses the heat conduction equation to constrain the prediction process of the physical constraint Transformer model. It also utilizes a sliding window mechanism to update the model weights every 15 minutes, enabling the model to adapt in real time to changes in power plant operating conditions. It also builds an expert rule library containing 200 thermodynamic criteria. When the model prediction conflicts with a thermodynamic criterion, the system immediately triggers a double verification mechanism to automatically call an independent physical model for recalculation. Combined with multi-source authentication by an expert team, this ensures that the prediction results conform to the actual evolution of the fire. Furthermore, the model's online learning mechanism regularly updates the weights via a sliding window, adapting in real time to changes in the complex operating conditions of the power plant, further enhancing the system's robustness in dynamic environments.

[0053] 5. This invention adopts a preemptive scheduling protocol, dynamically allocates resources through a dual-factor arbitration strategy of risk and timeliness, combines redundant channel design with high-fault-tolerant paths, and dynamically adjusts the PID parameters of the sprinkler system water pressure and robot path tracking. The system can still ensure the continuity of emergency response in extreme environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a system architecture diagram of the present invention;

[0055] Figure 2 It is a dynamic planning path flow chart of the present invention;

[0056] Figure 3 It is a detailed flow chart of the preemptive scheduling protocol of the present invention;

[0057] Figure 4 It is a schematic diagram of multi-physics field coupling of the present invention;

[0058] Figure 5 This is the architecture diagram of the regional collaborative disposal system of the present invention. DETAILED DESCRIPTION

[0059] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0060] Example 1

[0061] refer to Figure 1 、 Figure 2 and Figure 3 , an intelligent linkage disposal system for power plant fires, including a high-precision perception layer, a time series prediction engine, a game decision system and an execution control system.

[0062] Specifically, the high-precision perception layer consists of a distributed infrared temperature sensor array, a multi-plane gas monitoring unit, and a fiber optic deformation sensor.

[0063] Specifically, the infrared temperature sensor array uses infrared sensors distributed in a hexagonal grid topology, with a distance of 7m between adjacent nodes. A total of 150 sensors are deployed to cover the thermal field monitoring blind area, and the collected raw data are discrete temperature values; the multi-plane gas monitoring unit sets up three detection layers in the vertical direction of the steam turbine room, with the bottom layer height of 6m±3m, the middle layer height of 15m±4.5m, and the top layer height of 24m±6m. Each layer is equipped with 5 groups of CO, The sensor sampling frequency is 10Hz, and the raw data collected are CO, Gas concentration value; The optical fiber deformation sensor adopts FBG array structure and is deployed at the key nodes of the load-bearing beam. Each sensing unit contains 4 Bragg gratings. The raw data collected is the wavelength drift based on the FBG array. , real-time monitoring of structural deformation, structural deformation threshold .

[0064] Specifically, the high-precision perception layer deploys 10 edge computing nodes and uses a time-sensitive network to synchronize the raw data collected by sensors with a delay of ≤50ms. The pre-processed 12-dimensional feature vectors are transmitted to the timing prediction engine via industrial Ethernet.

[0065] Specifically, the edge computing node performs the following processing on the original data: spatial interpolation of discrete temperature values ​​using radial basis function neural network, and constructing precision Continuous temperature field model Generate continuous temperature field values, calculate spatial partial derivatives based on the continuous temperature field values ​​using Fourier's law, and derive the temperature gradient vector , and combined with the thermal conductivity of the material , through the formula Calculate heat flux At the same time, the time synchronization protocol is used to ensure the time synchronization of temperature data and derive the temperature change rate , the time window is 5s; the gas concentration value is denoised by Kalman filtering, and the gas diffusion gradient is calculated by Fick's law , and normalize the gas concentration values ​​after denoising of the three detection layers to generate vertical distribution features; through the Bragg grating formula The wavelength shift Convert to structural shape , using a 10-second sliding window to calculate the deformation rate Finally, the threshold value is compared to calculate the deviation between the structural deformation variable and the safety threshold value to generate risk warning characteristics; based on the continuous temperature field model In the gradient distribution of the spatial dimension and the dynamic evolution of the time dimension, the moving speed of the hot zone is extracted through the moving target detection algorithm , combined with CO concentration fluctuations to calculate the crowd flow index , move the hot zone at a speed of and crowd mobility index Dynamic characteristics of personnel.

[0066] Specifically, the temperature, gas concentration, deformation rate, and personnel dynamic characteristics are normalized to the [-1, 1] interval through the z-score to form a 12-dimensional feature vector.

[0067] Specifically, the time series prediction engine receives a 12-dimensional feature vector and uses a physical constraint Transformer model to extract features through a 6-layer encoder. Each encoder layer has 4 attention heads, and the 2nd and 4th layers are embedded through the heat conduction equation. The constructed regularizer.

[0068] Specifically, the pre-training phase of the physical constraint Transformer model uses a power plant fire dataset containing 500,000 samples. During transfer learning, the encoder parameters of the first three layers are frozen, and only the last three layers and the physical constraint module are fine-tuned.

[0069] Specifically, the loss function of the physical constraint Transformer model is ,in .

[0070] Specifically, the physical constraint Transformer model also adopts a sliding window online learning mechanism to update the model weights every 15 minutes. When it is detected that the working condition changes exceed the threshold, the model update is triggered in advance and a probability map of the fire evolution in the next 5 minutes is output.

[0071] Specifically, the time series prediction engine has built an expert rule library containing more than 200 thermodynamic criteria to verify the model output. Once the model prediction conflicts with the physical rules, the double verification mechanism is immediately activated.

[0072] Specifically, the dual verification mechanism is that when the fire evolution probability map output by the physical constraint Transformer model conflicts with more than 200 thermodynamic criteria in the expert rule base, including physical rules such as the heat conduction equation and the gas diffusion law, the system first calls another set of physical models independent of the original physical constraint Transformer model, inputs the same 12-dimensional feature vector for recalculation, and generates comparative fire prediction results. Secondly, it triggers the manual intervention interface and pushes the model prediction results and the dual physical model calculation results to the power plant safety monitoring center synchronously. The expert team conducts multi-source data cross-validation and finally determines the credibility of the prediction results by analyzing the consistency of the spatiotemporal distribution of temperature field, gas concentration field, and deformation rate.

[0073] Specifically, the game decision system includes a dynamic path planning module and a hierarchical equipment linkage controller.

[0074] Specifically, the dynamic path planning module uses a 15-second cycle, combines the environmental parameters and personnel dynamic characteristics in the fire evolution probability map, and generates three optimal evacuation paths through a multi-objective reinforcement learning algorithm: the shortest time path focuses on For fast-moving people, the path length optimization weight accounts for 60%, allowing crossing The medium-risk hotspot; the lowest-risk path targets Slow moving crowds, avoiding all Area, give priority to passing through equipment supply points with a distance of ≤15m; high fault tolerance path is suitable for In congested scenarios, the path contains ≥2 backup channels, allowing a maximum detour distance of 1.5 times the shortest path to ensure evacuation redundancy under extreme conditions.

[0075] Specifically, the specific implementation method of the multi-objective reinforcement learning algorithm to generate three optimal evacuation paths is as follows.

[0076] Specifically, the dynamic path planning module first extracts environmental parameters based on the fire evolution probability map, including temperature risk field ,visibility , Probability of passing through the escape route , and the hot zone movement speed in the personnel dynamic characteristics and crowd mobility index .

[0077] Specifically, since dynamic fire scenes have state uncertainty and real-time changes, the multi-objective reinforcement learning algorithm uses the Markov decision process to model them, abstracting environmental parameters and personnel dynamic characteristics into the state space of MDP, and defining operations such as path selection and direction adjustment as the action space, thereby constructing a dynamic mapping relationship of "state-action-reward".

[0078] Specifically, based on this modeling, the multi-objective reinforcement learning algorithm performs the following steps in each cycle: first, the real-time temperature field, gas concentration field, structural safety level, crowd distribution and other parameters are encoded into a state vector, and then based on the preset reward function, the 6-layer fully connected network parameters of the strategy function are optimized through the temporal difference algorithm to maximize the cumulative reward, and finally output three optimal paths adapted to different scenarios.

[0079] Specifically, the reward function is the shortest time path reward is negatively correlated with the path length, and the penalty is negatively correlated with the path length. High temperature area contact time, minimum risk path reward and avoidance The degree of high-risk area is positively correlated, the penalty is related to the amount of exposure to toxic gas, the reward for high-tolerance paths is positively correlated to the number of backup channels, and the penalty is related to the proportion of the detour distance exceeding the shortest path.

[0080] Specifically, the hierarchical equipment linkage controller has a built-in three-level response mechanism, which generates equipment control instructions based on the fire evolution probability map and 12-dimensional feature vector.

[0081] Specifically, the control logic for triggering the three-level response mechanism and generating device control instructions is as follows:

[0082] Specifically, in the first-level response, the hierarchical equipment linkage controller analyzes the fire evolution probability map output by the time series prediction engine in real time, and extracts the probability of an open flame occurring within the next minute. ,when And the duration that the fire probability exceeds the threshold continuously When the edge computing node is used, the following instructions are executed: control instructions are sent to the power monitoring system in the actuator through the edge computing node to shut down the power supply of non-critical areas, including the non-emergency circuit of the lighting system, to reduce the risk of electrical short circuit; the sound and light alarm system in the actuator is driven to issue intermittent alarms, and voice guidance is played through the fire alarm broadcast; the hot zone movement speed based on the 12-dimensional feature vector is detected. ,like The evacuation indicator light is controlled by the execution control system to indicate the temporary evacuation route by changing the flashing frequency.

[0083] Specifically, the trigger condition for the secondary response is the temperature change rate in the 12-dimensional feature vector or CO concentration , and the diffusion rate of high temperature area in the fire evolution probability diagram At this time, the grading equipment linkage controller executes the following instructions: send instructions to the intelligent smoke prevention and exhaust device, adjust the fan operation state to the medium speed gear, and give priority to eliminating the middle detection layer The fire door linkage module of the game decision system is controlled through the time-sensitive network to automatically lower the fire doors of the three adjacent fire partitions in the fire area, and mark the isolation area on the map module of the game decision system; according to the multi-plane gas monitoring unit Gas concentration value, if A cut-off command is sent to the fuel supply system of the execution control system to close the electromagnetic valve to block the hydrogen delivery.

[0084] Specifically, the triggering of the three-level response must simultaneously meet the highest temperature in the fire evolution probability map. , deformation rate And the population mobility index , the hierarchical equipment linkage controller sends the highest priority equipment control instructions through the time-sensitive network, including: sending partition control instructions to the variable sprinkler system, according to the temperature gradient Dynamically adjust the nozzle flow, including The regional nozzle pressure is increased to 1.2 times the rated value, forming a continuous water curtain isolation zone; the fire-fighting robot cluster is dispatched, guided by the high-tolerance path generated by the dynamic path planning module, combined with the structural safety level S of the fiber optic deformation sensor to avoid high-risk areas, and the fire extinguishing agent is sprayed after reaching the fire source; the emergency lighting system in the actuator is activated to force switching to the backup power supply to ensure that the illumination of the evacuation channel is ≥10lx, and the coordinates of the non-evacuated personnel are marked in real time based on multi-sensor data fusion technology, and coordinated rescue instructions are sent to the nearest fire-fighting robot.

[0085] Specifically, the execution control system receives the optimal evacuation path and equipment control instructions through the time-sensitive network to drive the actuators including the fire robot cluster, variable sprinkler system, and intelligent smoke prevention and exhaust device to execute the equipment control instructions, adopts the preemptive scheduling protocol to coordinate the actions of multiple devices, and provides real-time feedback on the equipment status to the hierarchical equipment linkage controller.

[0086] Specifically, the preemptive scheduling protocol builds a dynamic priority allocation mechanism based on risk level and task urgency.

[0087] Specifically, the risk level is divided into three levels: low, medium, and high based on typical fire scenarios in power plants. The digital twin model is used to evaluate the parameter thresholds including temperature, gas concentration, and structural deformation, and the weight coefficient is optimized through 2,000 simulation tests. , particle swarm optimization algorithm is used to iteratively optimize the weight coefficient , and finally determine the first-level priority corresponding to the high-risk scenario , Level 2 priority corresponds to medium-risk scenarios , Level 3 priority corresponds to low-risk scenarios .

[0088] Specifically, the dynamic priority allocation mechanism includes a three-level priority classification system: the first level priority, the fire extinguishing instruction of the fire robot cluster, is used for open fire fighting or structural safety level In high-risk scenarios, it can interrupt other non-critical tasks and monopolize communication bandwidth to ensure rapid arrival at the fire source; secondary priority, partition control instructions for variable sprinkler systems, when the temperature gradient vector or When the fire is in the vicinity of the fire source, its priority is automatically raised to the first level to ensure the rapid start of the water curtain isolation around the fire source; the third level priority, the ventilation adjustment instruction of the intelligent smoke exhaust device has the lowest default priority, but in the case of visibility And the population mobility index When the system is in operation, based on the multi-sensor data fusion results including the heat zone density of the infrared temperature sensor array, its priority is automatically raised to level two to ensure air circulation in the evacuation channel.

[0089] Specifically, the dynamic priority allocation mechanism also includes a dynamic preemption mechanism, which implements real-time scheduling through edge computing nodes and calculates the real-time risk weight of each executor every 5 seconds. , when the real-time risk weight of an actuator Value exceeds threshold Emergency preemption is triggered.

[0090] Specifically, to address the resource competition problem among multiple executors, the preemptive scheduling protocol adopts a dual-factor arbitration strategy of risk and timeliness. When multiple executors request the same resource, the arbitration score formula is used: Conduct an assessment, where is the current risk value of the device, The person with the highest score for the remaining time on the task gains control of the resources.

[0091] Specifically, the hierarchical device linkage controller dynamically adjusts PID parameters or path switching based on the device status deviation by executing the device status feedback in real time from the control system. The specific implementation method is as follows:

[0092] Specifically, when the execution control system feeds back the pipe water pressure value of the variable sprinkler system and target value Deviation exceeds When the controller extracts the deviation signal and rate of change , automatically adjust the proportional coefficient through PID algorithm :like and , when the water pressure continues to be lower than the target value, Increase the pressure in steps of 0.1 to 1.2 to speed up the boost response; if and , when the water pressure is close to stable, Decrease to 0.6 to suppress overshoot; at the same time, when the water pressure deviation continues to fail to converge for more than 30 seconds, the standby pump start command is automatically triggered.

[0093] Specifically, for the real-time position feedback of the fire-fighting robot cluster, if it is detected that a robot deviates from the optimal path generated by the dynamic path planning module by more than 0.5m, and the optical fiber deformation sensor does not feedback the structural deformation warning of the area, the hierarchical equipment linkage controller immediately selects the suboptimal path with the shortest Euclidean distance to the current position from the three pre-generated optimal paths, sends a path replanning instruction to the robot, and adjusts the PID parameters of its drive motor at the same time: increase the differential coefficient To enhance path tracking rigidity, ensure return to the new path within 15 seconds, and path switching delay ≤ 1.5 seconds.

[0094] Specifically, when the intelligent smoke exhaust device reports that the fan speed fluctuation exceeds the rated value When the grading equipment linkage controller is combined with the CO concentration gradient of the multi-plane gas monitoring unit Make a joint judgment: If the CO concentration gradient If the CO concentration gradient exceeds 0.3ppm / m, it indicates abnormal airflow disturbance and is judged to be caused by pipeline blockage, resulting in abnormal speed. A shutdown command is immediately sent to trigger the start of the backup fan, and a mechanical clearance request is sent to the execution control system. Normal, recalculate the integral coefficient through the PID parameter self-tuning algorithm , suppress the speed fluctuation to within the range.

[0095] Specifically, during the process of dynamically adjusting the PID parameters, the hierarchical device linkage controller records the feedback device status in real time and uploads it to the timing prediction engine through the edge computing node as a training sample for the physical constraint Transformer model to optimize the prediction accuracy of future control instructions.

[0096] Example 2

[0097] refer to Figure 1 、 Figure 3 and Figure 4 Based on the basic system architecture of Example 1, this embodiment implements more accurate fire prevention and control and coordinated disposal through the following technical improvements in response to the coupling of multiple physical fields under complex working conditions of power plants.

[0098] Specifically, in order to fully capture the multi-dimensional signals during the fire incubation period, the high-precision sensing layer adds distributed strain sensors and dust concentration sensors in the cable interlayer to collect structural strain in real time. and dust concentration .

[0099] Specifically, the edge computing node introduces a multi-physics field coupling algorithm to fuse data through thermal-structure coupling and airflow-dust coupling.

[0100] Specifically, thermal structure coupling is based on the temperature gradient vector of the infrared temperature sensor array , through the thermal expansion coefficient formula Calculate the thermal deformation, where the thermal expansion coefficient of the material , L is the length of the structural member, thermal deformation The structural deformation variable ε of the optical fiber deformation sensor is weightedly fused with the weight ratio of 0.7:0.3, and the judgment formula of the modified structural safety level S is: ,in, The safety threshold is used to improve the accuracy of structural risk assessment through the direct correlation between gradient and deformation.

[0101] Specifically, the airflow dust coupling is combined with the CO concentration gradient , crowd mobility index and dust concentration , derive the airflow disturbance velocity through the fluid diffusion model , the calculation formula for corrected visibility is: , the effect of enhancement gradient parameters on gas diffusion.

[0102] Specifically, the edge node finally generates a 15-dimensional feature vector containing temperature, gas, deformation, strain, and dust.

[0103] Specifically, the physical constraint Transformer model embeds multi-physics field regularization terms in the 2nd, 4th, and 6th layer encoders, including the heat conduction equation, gas diffusion equation, and structural mechanics equation. The loss function is expanded to: in, is the gas flux, is the diffusion coefficient, is stress, is the elastic modulus.

[0104] Specifically, a convolutional neural network-based operating condition classifier is introduced to identify the power plant operating modes, including normal, maintenance, and start-up and shutdown, in real time. When the "maintenance condition" is detected, the physical constraint Transformer model is automatically adjusted to increase the fire probability threshold to 45%, and the regularization terms other than the heat conduction equation are frozen to avoid false alarms caused by heat sources such as welding.

[0105] Specifically, the dynamic path planning module adds dust risk field to the three-dimensional thermal risk assessment system. ,in ,when Or when the structural safety level S ≥ 4, the path planning is forced to avoid this area.

[0106] Specifically, for multi-field coupling scenarios, composite risk paths are generated, with priority given to high ventilation efficiency and structural safety level. And the structural strain channel.

[0107] Specifically, the grading equipment linkage controller adds a dust suppression instruction in the secondary response: when the dust concentration and Yes, the high-pressure spray system is started simultaneously to reduce the dust concentration and the diffusion rate of combustible gas by droplet adsorption; the structural safety warning is introduced in the third-level response: if the strain sensor detects the structural strain of the load-bearing beam And the deformation rate\ , triggering the emergency channel to lock, prohibiting personnel from entering and dispatching robots to evacuate quickly.

[0108] Specifically, the preemptive scheduling protocol of the execution control system introduces the physical field risk weight ,in for Concentration, when When an emergency occurs, the priority of the fire-fighting robot cluster jumps to level one, and it can interrupt the non-emergency tasks of the smoke prevention and exhaust system.

[0109] Specifically, for the competition for water resources between the firefighting machine cluster and the variable sprinkler system, a risk and time arbitration formula is adopted: Ensure that firefighting needs in high-risk areas are met first.

[0110] Specifically, the PID parameter adjustment embedded in the multi-physics compensation item includes: when the variable sprinkler system water pressure is adjusted, if it is detected , automatically scale the coefficient The gain is increased by 20%, which speeds up the water curtain isolation speed; when the fire-fighting robot passes through the abnormal strain area, the differential coefficient according to Dynamically enhance the path avoidance capability in structural deformation areas.

[0111] Example 3

[0112] refer to Figure 1 and Figure 3 Based on the system architecture of Example 1 and Example 2, this embodiment targets the special requirements of the nuclear power plant radiation environment for fire prevention and control through hardware radiation resistance modification, data model adaptation and emergency strategy adjustment.

[0113] Specifically, the high-precision sensing layer enhances the sensor's radiation resistance. The sensors of the infrared temperature sensor array and the multi-plane gas detection unit are all packaged with lead shielding. The lead shielding layer is 5mm thick, and the sensor electronic components are radiation-resistant models to ensure that the temperature collection error in the radiation field is ≤±0.5℃ and the gas concentration error is ≤±3%. The FBG array of the optical fiber deformation sensor uses radiation-hardened optical fiber, and the Bragg grating period stability is improved to ±0.01nm / year to avoid wavelength drift errors caused by radiation.

[0114] Specifically, the edge computing node uses FPGA combined with radiation-resistant DSP architecture, and the storage medium is replaced with magnetoresistive random access memory to ensure the anti-interference capability of data processing and transmission. The mean time between failures (MTBF) of a single node is ≥ 100,000 hours.

[0115] Specifically, the radiation dose rate is added to the 15-dimensional feature vector , real-time monitoring of environmental radiation dose rate through Geiger counter , and establish a radiation and thermal coupling model: ,in, is the corrected temperature, is the radiation heat effect coefficient, and t is the cumulative radiation time, which is used to eliminate the interference of radiation temperature rise on fire prediction.

[0116] Specifically, the radiation diffusion equation is introduced into the regularization term of the physical constraint Transformer model, and the loss function is expanded to: , ensuring that the fire evolution probability map output by the model eliminates the misjudgment effect of radiation temperature rise, and improves the prediction accuracy of strong radiation areas such as nuclear islands, with an error of ≤8%.

[0117] Specifically, the dynamic path planning module adds radiation risk field ,when When performing tasks, the path is forced to avoid high-dose radiation areas, and lead-protected channels or remote-controlled robot clusters are given priority.

[0118] Specifically, the triggering conditions of the three-level response mechanism and the generation of equipment control instructions are enhanced. The second-level response: if the environmental radiation dose rate And the radiation exceeds the threshold for a period of time After 30 minutes, the firefighting robot cluster will automatically switch to remote control mode, and the real-time images will be transmitted back through the explosion-proof camera to reduce the risk of human intervention; Level 3 response: combined with the environmental radiation dose rate And the structural deformation , triggering an emergency evacuation order for the entire plant, and at the same time activating the radiation shielding doors to isolate the fire area, blocking the combined risks of radiation leakage and fire spread.

[0119] Specifically, the fire-fighting robot in the actuator fire-fighting robot cluster uses a tungsten alloy radiation-resistant shell, the drive motor is upgraded to a radiation-resistant brushless DC motor, and is equipped with a redundant navigation system including vision, inertial navigation, and RFID landmarks to ensure continuous operation for ≥2 hours in smoke and radiation environments, with a path tracking error of ≤0.8m.

[0120] Specifically, the priority of radiation-resistant equipment in the preemptive scheduling protocol is automatically upgraded to the second-level priority, and the risk weight formula is optimized as follows: , when the radiation dose rate When the fire is extinguished, the firefighting task priority of the fire-fighting robot cluster jumps to the first level priority, which can interrupt the control instructions of non-radiation-resistant equipment to ensure priority allocation of resources in the early stage of the fire.

[0121] Specifically, the PID control parameters are compensated when the sprinkler system detects When the scale factor is automatically Gain increased by 15%, speeding up the formation of water curtain; differential coefficient of robot path tracking According to the dynamic adjustment of radiation dose, the dynamic adjustment formula is: .

[0122] Example 4

[0123] refer to Figure 1 、 Figure 3 and Figure 5 Based on the system architecture of Example 1, Example 2 and Example 3, this embodiment aims at the fire prevention and control needs of large power plant clusters or regionalized fires by building a regional collaborative disposal system across plant areas.

[0124] Specifically, edge computing node clusters are deployed in each power plant in the region. The high-precision perception layer of each power plant is interconnected with the edge computing node cluster through the 5G industrial private network, and the feature vectors generated by the high-precision perception layer after preprocessing are synchronized in real time to form a distributed perception network with a coverage radius of 50 kilometers.

[0125] Specifically, the edge node of each power plant will encrypt and upload the preprocessed feature vector to the cloud data center, and construct a regional fire risk heat map through spatiotemporal alignment and data normalization.

[0126] Specifically, the regional fire risk heat map is based on the real-time data of the power plant combined with the thermal field data of adjacent power plants to calculate the regional risk index. , dynamically identify fire spread trends and high-risk areas, including For fire risk, For radiation risks, It is the risk of population density.

[0127] Specifically, the dynamic path planning module uses a regional global optimization algorithm, combined with the connectivity of the evacuation channels and equipment status of each power plant, to generate an evacuation path across the plant area.

[0128] Specifically, when a fire occurs in a single power plant, the dynamic path planning module breaks through the plant boundaries and generates a cross-plant evacuation route based on the connectivity of the evacuation channels and the distribution of fire-fighting equipment of each power plant.

[0129] Specifically, the triggering conditions of the three-level response mechanism and the generation of equipment control instructions are enhanced. The first-level response is a regional early warning. When the probability of fire in a power plant is If a similar heat source anomaly occurs in an adjacent power plant, the cloud triggers a regional warning, uniformly adjusts the sound and light alarm frequency and evacuation indicator light direction of each power plant, and guides personnel to gather in the cross-plant corridor; the third-level response is cross-plant support. If a single power plant fire gets out of control And the population mobility index The cloud forcibly takes over the actuators of adjacent power plants, dispatches idle fire-fighting robot clusters for cross-plant support, and simultaneously links with the municipal fire-fighting system to open regional fire-fighting water sources.

[0130] Specifically, the data from each power plant is collaboratively trained on the cloud through federated learning to train the physical constraint Transformer model, and the model parameter updates are shared while protecting privacy, thereby improving the overall prediction accuracy of the region.

[0131] Specifically, virtual simulations are conducted regularly on the cloud to simulate regional disaster scenarios based on historical fire data, optimize resource scheduling strategies including the routes of firefighting robots and water allocation priorities, and generate standardized emergency response plans.

[0132] Specifically, the cloud-based operating console provides augmented reality visualization, displaying real-time fire spread simulation, personnel distribution, and equipment status of each power plant, supporting remote intervention by expert teams to adjust strategies.

[0133] Specifically, to ensure system reliability in extreme environments, a dual-link communication network combining 5G and satellite relay is built. When the factory network is damaged by fire, the edge nodes transmit key data including personnel coordinates and equipment status to the cloud via satellite. The communication delay is ≤200ms, ensuring the continuity of command issuance.

[0134] Specifically, a regional emergency energy pool will be established, and the energy storage equipment of each power plant will be connected to the cloud for management. When the backup power supply of a power plant is exhausted, the cloud will automatically dispatch energy support from adjacent power plants to ensure the continuous operation of key equipment including sprinkler systems and emergency lighting.

[0135] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

[0136] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. An intelligent linkage fire handling system for power plants, characterized in that: include: The high-precision perception layer consists of a distributed infrared temperature sensor array, multi-plane gas monitoring units, and fiber optic deformation sensors. The raw data collected by each sensor is preprocessed by edge computing nodes to generate a 12-dimensional feature vector consisting of temperature, gas concentration, deformation rate, and human dynamic characteristics within a continuous time window. The time series prediction engine uses a physically constrained Transformer model, takes a 12-dimensional feature vector as input, and outputs a spatially gridded fire evolution probability map, which is then synchronously transmitted to the game decision-making system. The game decision-making system includes a dynamic path planning module and a hierarchical equipment linkage controller. The dynamic path planning module uses a multi-objective reinforcement learning algorithm to generate three optimal evacuation paths based on the environmental parameters and personnel dynamic characteristics in the fire evolution probability map in a 15-second cycle. The hierarchical equipment linkage controller has a built-in three-level response mechanism that generates equipment control instructions based on the fire evolution probability map and 12-dimensional feature vectors. The execution control system receives the optimal evacuation path and equipment control instructions through the time-sensitive network to drive the actuators including the fire robot cluster, variable sprinkler system, and intelligent smoke prevention and exhaust device to execute the equipment control instructions. It uses a preemptive scheduling protocol to coordinate the actions of multiple devices and provides real-time feedback on the equipment status to the hierarchical equipment linkage controller for dynamic adjustment of PID parameters or path switching.

2. The intelligent linkage fire handling system for power plants according to claim 1, characterized in that: In the high-precision perception layer, the infrared temperature sensing array uses infrared temperature sensors and is deployed according to a hexagonal grid topology. The distance between adjacent nodes is 6m < d < 8m, and the original data collected is discrete temperature values. The multi-plane gas monitoring unit has three detection layers arranged vertically, and the installation height h is proportionally allocated based on the site height H. The height of the bottom detection layer , the height of the middle detection layer , the height of the top detection layer . The sensors configured in the multi-plane gas monitoring unit include CO sensors and sensors. The original data collected is the gas concentration values collected by the sensors at a sampling frequency of 10Hz. The fiber optic deformation sensor uses a FBG array structure, and each sensing unit contains 4 Bragg gratings. The original data collected is the wavelength drift .

3. The intelligent linkage fire handling system for power plants according to claim 1 or 2, characterized in that: The data preprocessing includes: For discrete temperature values, radial basis function neural network is used for spatial interpolation to build precision Continuous temperature field model Generate continuous temperature field values, calculate spatial partial derivatives based on the continuous temperature field values ​​using Fourier's law, and derive the temperature gradient vector , and combined with the material thermal conductivity k, through the formula Calculate the heat flux density q, and ensure the time synchronization of temperature data through the time synchronization protocol, and derive time series characteristics including temperature change rate ; For the gas concentration value, the airflow disturbance noise is suppressed by filtering technology, and the denoised CO, Concentration value: The gas diffusion gradient is calculated based on the gas concentration values ​​of the three detection layers using Fick's law, and the gas concentration values ​​of the three detection layers are normalized to a unified coordinate system to generate vertical concentration distribution characteristics; For wavelength shift Converting to structural deformation through Bragg grating formula ,in is the refractive index of light, is the grating period, and the deformation rate is calculated within the time window using the sliding window algorithm Evaluate the dynamic trend of the structure, and finally perform threshold comparison to calculate the deviation between the structural deformation and the safety threshold to generate risk warning features; At the same time, the dynamic characteristics of personnel are obtained through multi-sensor data fusion. First, based on the continuous temperature field model In the gradient distribution of the spatial dimension and the dynamic evolution of the time dimension, the moving trajectory and speed of the high-temperature area in the continuous temperature field are calculated through the moving target detection algorithm to extract the moving speed of the hot zone reflecting the movement of people. Then, the abnormal fluctuation of gas concentration value is analyzed by CO concentration value. When the CO concentration shows a sudden change of non-natural diffusion, the degree of disturbance of air flow by crowd flow is inferred based on the fluid mechanics model, and the crowd flow index is calculated. Finally, the parameters are fused to generate the hot zone moving speed and crowd mobility index Dynamic characteristics of personnel.

4. The intelligent linkage fire handling system for power plants according to claims 1 and 3, characterized in that: The 12-dimensional feature vector includes temperature, gas concentration, deformation rate and personnel dynamic characteristics, and the temperature feature includes continuous temperature field value, temperature gradient vector , heat flux density q, time series characteristics; the gas concentration characteristics include CO and Concentration value, gas diffusion gradient, vertical concentration distribution characteristics; the deformation rate characteristics include structural deformation , deformation rate , risk warning characteristics; the personnel dynamic characteristics include the speed of movement in hot areas and population mobility index The above temperature, gas concentration, deformation rate and personnel dynamic characteristics are normalized to the [-1,1] interval through z-score to form a 12-dimensional feature vector that meets the input requirements of the time series prediction engine.

5. The intelligent linkage fire handling system for power plants according to claim 1, characterized in that: The physical constraint Transformer model used by the time series prediction engine has the following features: The physical constraint Transformer model adopts a 6-layer encoder stack structure, with 4 independent attention heads configured on each layer. The 2nd and 4th layers introduce physical constraint modules, and the heat conduction equation is used to calculate the attention head. Construct a regularization term, where α is the material thermal diffusivity and Q is the heat source term; the loss function of the physical constraint Transformer model is ,in The L2 regularization term is determined by cross-validation optimization, with β=0.3 and γ=0.

1. The physical constraint Transformer model is pre-trained on a power plant fire dataset containing 500,000 samples for pre-training and transfer learning. During transfer learning, the encoder parameters of the first three layers are frozen, and only the last three layers and the physical constraint module are fine-tuned. The physical constraint Transformer model adopts a sliding window online learning mechanism, updating the weights every 15 minutes. When the operating condition change exceeds the threshold, the model update is triggered in advance.

6. The intelligent linkage fire handling system for power plants according to claim 1 or 5, characterized in that: To ensure the reliability of the output of the physical constraint Transformer model, the time series prediction engine constructed an expert rule library containing 200 thermodynamic criteria to verify the model output. When the fire evolution probability map output by the physical constraint Transformer model conflicts with the thermodynamic criteria in the expert rule library, a dual verification mechanism is immediately activated. The system first calls another set of physical models independent of the original physical constraint Transformer model, inputs the same 12-dimensional feature vector for recalculation, and generates comparative fire prediction results. Then, it triggers the manual intervention interface and synchronously pushes the model prediction results and the dual physical model calculation results to the power plant safety monitoring center. The expert team conducts multi-source data cross-validation and ultimately determines the credibility of the prediction results by analyzing the consistency of the spatiotemporal distribution of temperature fields, gas concentration fields, and deformation rates.

7. The intelligent linkage fire handling system for power plants according to claim 1 or 5, characterized in that: The dynamic path planning module obtains environmental parameters based on the fire evolution probability map, including the temperature gradient vector ,visibility , Probability of passing through the escape route Construct a three-dimensional thermal risk assessment system based on temperature gradient vector Defining the temperature risk field , The mapping relationship between CO concentration and visibility based on the Beer-Lambert law , Probability of passing through the escape route Assess environmental risks and combine them with personnel dynamic characteristics including movement speed in hot areas , crowd mobility index , through multi-objective reinforcement learning, three optimal paths are generated every 15 seconds, including the shortest time path, the lowest risk path, and the high fault tolerance path, among which the shortest time path focuses on For fast-moving people, the path length optimization weight accounts for 60%, allowing crossing The medium-risk hotspot; the lowest-risk path targets Slow moving crowds, avoiding all Area, give priority to passing through equipment supply points with a distance of ≤15m; high fault tolerance path is suitable for In congested scenarios, the path contains ≥2 backup channels, and the maximum detour distance allowed is 1.5 times the shortest path.

8. A power plant fire intelligent linkage handling system according to claim 1 or 7, characterized in that: The hierarchical equipment linkage controller constructs a three-level response mechanism based on the fire evolution probability map and the 12-dimensional feature vector, including: The first level response is And duration When the hot zone is moving, the following control instructions are generated: turn off the power supply of non-critical areas, start the sound and light alarm, and Dynamically adjust the evacuation indicator light, including the fire probability The probability value of the fire evolution probability diagram in the next 1 minute, the duration is the duration that the fire probability exceeds the threshold continuously; The secondary response is when the temperature change rate or CO concentration , and the diffusion rate of high temperature area in the fire evolution probability diagram When the fire door is closed, the following control instructions are generated: start the intelligent smoke exhaust device, adjust the fan operation state to the medium speed gear, lower the adjacent fire door, and monitor the gas flow in real time through the multi-plane gas detection unit. Gas concentration value, when Time cut off supply; The third-level response is when the highest temperature in the fire evolution probability map , deformation rate And the population mobility index When the temperature gradient is Dynamically adjust the flow rate of sprinklers, dispatch firefighting robot clusters to extinguish fires along high-tolerance paths, switch emergency lighting systems, and mark the coordinates of people who have not evacuated.

9. The intelligent linkage fire handling system for power plants according to claim 1 or 8, characterized in that: The execution control system adopts a preemptive scheduling protocol based on risk weights. Dynamically assign priorities when Emergency preemption is triggered when For multi-executor resource competition, the arbitration scoring formula is used Assessment, where is the current risk value of the device, The person with the highest score for the remaining time on the task gains control of the resources.

10. The system according to claim 1 or 9, characterized in that The execution control system feeds back the device status to the hierarchical device linkage controller in real time via a time-sensitive network. The hierarchical device linkage controller dynamically adjusts PID parameters or path switching based on the fed-back device status, specifically including: When the execution control system feeds back the pipe water pressure value of the variable sprinkler system and target value When the deviation exceeds ±10%, the grading equipment linkage controller extracts the deviation signal and rate of change , automatically adjust the proportional coefficient through PID algorithm :If the water pressure continues to be lower than the target value and , Increase the value by 0.1 to 1.

2. If the water pressure is close to stable, Decrease to 0.6 to suppress overshoot. If the water pressure deviation does not converge for more than 30 seconds, the backup pump start command is triggered; For the real-time position feedback of the firefighting robot cluster, if it is detected that the robot deviates from the optimal path by more than 0.5m and there is no structural deformation warning, the hierarchical equipment linkage controller selects the suboptimal path from the three pre-generated optimal paths and replans, and adjusts the differential coefficient of the drive motor at the same time. , ensuring that the new path is returned within 15 seconds, and the path switching delay is ≤ 1.5 seconds; When the intelligent smoke exhaust device feedback fan speed fluctuation exceeds the rated value ±15%, the grading equipment linkage controller combines the CO concentration gradient to make a judgment: if the CO concentration gradient is > 0.3ppm / m, the pipeline is determined to be blocked and the backup fan is switched; if it is normal, the integral coefficient is recalculated through the PID parameter self-tuning algorithm , suppressing speed fluctuations within the range of ±5%; the hierarchical equipment linkage controller records the equipment status in real time and uploads it to the timing prediction engine through the edge computing node as a training sample for the physical constraint Transformer model. The physical constraint Transformer model optimizes the prediction accuracy of control instructions through an online learning mechanism.

Images