A fire warning method and system based on computing power allocation

Abstract

The application discloses a fire warning method and system based on computing power distribution, and belongs to the technical field of fire warning. The method aims to solve the technical problems of the existing technology, such as the incapability of traditional warning to predict fire situation, the incapability of offline computational fluid dynamics simulation to respond to real-time data, and the waste of computing power. The technical scheme comprises the following steps: acquiring a sensor data stream in real time, generating an initial fire scene in response to an abnormal signal; simulating an active node scene in parallel by using a computational fluid dynamics model, and expanding a scene tree based on a physical rule set; periodically comparing simulation and real-time data, updating node probability by using Bayes, pruning invalid nodes and recycling computing power, determining priority based on node probability and simulation lag degree, and dynamically distributing computing power; repeating iteration, and outputting a warning based on the simulation result of a high-probability node. The application realizes online application of computational fluid dynamics simulation, and improves computing power utilization rate and warning accuracy.

Description

Technical Field

[0001] This invention relates to the field of fire early warning technology, and more specifically, to a fire early warning method and system based on computing power allocation. Background Technology

[0002] Fire is a major risk source threatening the safe operation of high-risk industrial facilities. Traditional fire early warning relies on threshold alarms from point sensors, which are slow to respond and cannot predict the development of the situation. Fire simulation based on computational fluid dynamics (CFD) (such as using FDS software) can reproduce the physical process of fire with high fidelity and has become an important analytical tool.

[0003] However, using CFD simulation for real-time early warning faces fundamental bottlenecks. For example, Chinese patent CN111814337A discloses a method for fire risk analysis in nuclear power plant compartments. Although this solution integrates CFD simulation, its essence is to use CFD as an offline, post-hoc engineering design and analysis tool. The technical solution involves pre-selecting several typical fire scenarios during the facility design or safety review phase, and using CFD software to simulate and calculate these static, pre-defined scenarios to assess risks and optimize fire protection design. This solution represents the current mainstream application mode of CFD in the field of fire safety.

[0004] The existing technical solution suffers from two significant, interconnected flaws that render it unsuitable for real-time dynamic early warning: First, the computation mode is entirely offline and open-loop. All simulated scenarios are pre-defined statically, and the simulation process cannot access or respond to the real-time changes in sensor data streams during facility operation, resulting in a severe disconnect between the calculation results and the actual dynamic fire situation. Second, the utilization rate of computing resources is extremely low and rigid. Whether it is a single-scenario simulation or multi-scenario parallel computing, computing resources are fixedly allocated at the start of the task and remain so until the simulation is completed. Even if the assumptions of a certain simulated scenario are quickly proven to be inconsistent with reality, the system cannot terminate the invalid calculation midway, causing valuable computing resources (such as high-performance computing clusters) to be continuously wasted on incorrect extrapolation paths. The fundamental problem is that the existing technology lacks a mechanism that can dynamically guide the high-cost CFD computation process based on real-time sensing data and intelligently schedule computing resources, making CFD simulations unsuitable for real-time emergency decision-making due to their excessive time consumption. Summary of the Invention

[0005] The purpose of this invention is to provide a fire early warning method and system based on computing power allocation, so as to solve the technical problems in the prior art that traditional early warning cannot predict the fire situation, offline CFD simulation cannot respond to real-time data and wastes computing power.

[0006] To solve the above-mentioned technical problems, the present invention provides a fire early warning method based on computing power allocation, comprising the following steps: S1. Real-time acquisition of sensor data streams in the fire monitoring area; S2. In response to abnormal signals in the sensor data stream, generate at least one initial fire scenario based on the rapid diagnostic model, which serves as the root node of the dynamic probabilistic fire scenario tree. S3. Using a computational fluid dynamics model, perform parallel simulation calculations on the fire scenarios corresponding to each active node in the dynamic probabilistic fire scenario tree; S4. Based on the simulated field data at a specific time step generated by parallel simulation calculation, and based on a preset set of physical rules, deduce and generate sub-scene nodes of the currently active nodes to expand the dynamic probability fire scene tree. S5. Periodically compare the predicted data simulated by the computational fluid dynamics model corresponding to each active node with the real-time sensor data, and dynamically update the real-time probability value of each active node based on the comparison results. S6. Based on the preset elimination conditions, prune active nodes and their descendant nodes whose real-time probability values ​​are lower than the first threshold or whose deviation between predicted data and sensor data is consistently higher than the second threshold, and terminate the computational fluid dynamics model simulation task corresponding to them. S7. Based on the real-time probability value of each active node and the degree to which the simulation progress of the corresponding computational fluid dynamics model lags behind the real time, determine the scheduling priority of each active node, and dynamically allocate computing resources to the active nodes in the dynamic probability fire scenario tree according to the scheduling priority; wherein, the computing resources reclaimed from the pruned nodes are redistributed to active nodes with higher scheduling priority. S8. Repeat steps S3 to S7, and generate and output fire development prediction and early warning information based on the simulation results of the fluid dynamics model corresponding to the active nodes whose current real-time probability value is higher than the third threshold.

[0007] As a further improvement to this technical solution, in step S2, the rapid diagnostic model is a machine learning-based classifier or a rule-based expert system; its input includes at least the sensor type, location, signal amplitude, and rate of change of the sensor that triggers the abnormal signal, and its output is at least one fire scenario hypothesis, which includes the fire source location, initial fire source power, and fire growth model parameters.

[0008] As a further improvement to this technical solution, in step S4, the physical rule set includes one or more preset physical evolution rules. The execution conditions of the physical evolution rule are based on intermediate result parameters generated by the computational fluid dynamics model simulation. The execution action of the physical evolution rule is to generate one or more new sub-scene nodes. The intermediate result parameters include at least one of predicted temperature, smoke concentration, and thermal radiation flux. The generation of the sub-scene node includes modifying the boundary conditions or fire source terms of the fire scene corresponding to the currently active node, which serves as the starting point for the deduction.

[0009] As a further improvement to this technical solution, in step S5, the real-time probability value of each active node is dynamically updated based on the comparison results, which is implemented using a Bayesian update method, specifically including: The likelihood of each active node is calculated based on the degree of agreement between the predicted data from the computational fluid dynamics model and the measured data from the corresponding position sensor; the higher the degree of agreement, the higher the likelihood. Multiply the current real-time probability value of each active node by its corresponding likelihood to obtain the updated probability weight; The updated probability weights of all active nodes are normalized so that the sum of the probability weights of all active nodes is 1, thus obtaining the new real-time probability value of each node.

[0010] As a further improvement to this technical solution, in step S6, the elimination conditions include: Condition 1: The real-time probability value of the node is lower than the first threshold for N consecutive update cycles; Condition 2: The error between the key parameters simulated and predicted by the nodal computational fluid dynamics model and the corresponding measured data from the sensors exceeds the second threshold for M consecutive update cycles; The pruning operation is triggered when either condition one or condition two is met, where N and M are preset positive integers.

[0011] As a further improvement to this technical solution, in step S7, the scheduling priority of each active node is determined based on its real-time probability value and the degree to which the progress simulated by the computational fluid dynamics model lags behind the real time. The higher the real-time probability value of a node, and the greater the lag between its computational fluid dynamics model simulation progress and real time, the higher its scheduling priority.

[0012] As a further improvement to this technical solution, the determination of the scheduling priority of each active node is also based on a key flag bit: When the fire scenario represented by a node involves an impact on predefined safety-critical equipment or areas, the critical flag of that node is activated. Nodes whose critical flags are activated will have their scheduling priority increased under the same conditions.

[0013] As a further improvement to this technical solution, the specific steps for dynamically allocating computing resources in step S7 include: S71. Obtain the scheduling priority of all currently active nodes and sort them according to priority. S72. Divide the available computational fluid dynamics model computational core resources into multiple computational resource units according to a preset ratio; S73. Allocate the first-ranked active node with the first proportion of computing resource units, and allocate the remaining computing resource units proportionally according to the scheduling priority of each node. S74. Periodically or upon receiving a node pruning notification, re-execute steps S71 to S73.

[0014] As a further improvement to this technical solution, in step S3, the computational accuracy of the computational fluid dynamics model of different active nodes is dynamically adjusted according to their scheduling priority; wherein, nodes with scheduling priority higher than the fourth threshold adopt the first computational accuracy, and nodes with scheduling priority lower than the fourth threshold adopt the second computational accuracy lower than the first computational accuracy.

[0015] A fire early warning system based on computing power allocation, wherein the fire early warning system based on computing power allocation is used to implement the above-mentioned fire early warning method based on computing power allocation, comprising: The data acquisition module is configured to acquire sensor data streams from the fire monitoring area in real time. The scene management engine, connected to the data acquisition module, is used to respond to abnormal signals in the sensor data stream, generate at least one initial fire scene based on a rapid diagnostic model, and serve as the root node of the dynamic probabilistic fire scene tree; periodically compare the predicted data simulated by the computational fluid dynamics model corresponding to each active node with the real-time acquired sensor data, and dynamically update the real-time probability value of each active node based on the comparison results; according to preset elimination conditions, prune active nodes and their descendant nodes whose real-time probability values ​​are lower than a first threshold or whose deviation between the predicted data and the sensor data is consistently higher than a second threshold. A computational scheduling engine, connected to the scene management engine, controls the parallel computing cluster to perform parallel simulation calculations of the fire scenarios corresponding to each active node in the dynamic probabilistic fire scenario tree using a computational fluid dynamics model, and feeds back the simulation field data at a specific time step to the scene management engine; based on the simulation field data and a preset set of physical rules, it infers and generates sub-scene nodes of the currently active nodes to expand the dynamic probabilistic fire scenario tree; based on the real-time probability value of each active node obtained from the scene management engine and the degree to which the computational fluid dynamics model simulation progress lags behind the real time, it determines the scheduling priority of each active node; according to the scheduling priority, it dynamically allocates the computing resources of the parallel computing cluster to the active nodes in the dynamic probabilistic fire scenario tree, and in response to the pruning command of the scene management engine, terminates the computing task of the corresponding node and reallocates the released resources; The early warning output module is connected to the scene management engine. Based on the computational fluid dynamics model simulation results of active nodes whose current real-time probability value is higher than the third threshold, it generates and outputs fire development prediction and early warning information.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By constructing a closed loop of real-time data-driven, probabilistic scenario tree dynamic evolution, and adaptive scheduling of computing resources, the originally offline and time-consuming CFD simulation was successfully applied to online real-time early warning. This forces computing resources to always focus on the most likely and most urgent fire evolution paths and eliminates invalid paths in real time. As a result, physically based situation predictions are obtained during the emergency window period of the initial fire. At the same time, it avoids the waste of computing resources running in the wrong direction and improves the utilization efficiency of high-value computing clusters.

[0017] 2. By introducing a probabilistic dynamic correction mechanism based on Bayesian updates, the system can learn and correct itself as the fire situation develops, enhancing the adaptability and reliability of the early warning.

[0018] 3. By considering a composite priority scheduling strategy that takes into account both simulation progress lag and scenario criticality, the timeliness of emergency response has been further optimized.

[0019] 4. By applying differentiated computational precision to tasks of different priorities, fine-grained control over the trade-off between computational speed and precision is achieved under a fixed computing power constraint. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall method of the present invention; Figure 2 This is an overall system block diagram of the present invention. Detailed Implementation

[0021] Example 1 Currently, traditional fire early warning relies on threshold alarms from point sensors, which are slow to respond and cannot predict the development of the situation. Fire simulation based on computational fluid dynamics (CFD) (such as using FDS software) can reproduce the physical process of fire with high fidelity and has become an important analysis tool. However, existing fire analysis using computational fluid dynamics (CFD) has problems such as completely offline and open-loop calculation mode, as well as extremely low and rigid utilization of computing resources. In view of this, please refer to Figure 1 As shown, one of the objectives of this invention is to provide a fire early warning method based on computing power allocation. This embodiment describes the application of this method in fire early warning of auxiliary compartments in nuclear power plant reactor buildings. The implementation process of this method is explained in detail below with specific steps: Considering that real-time and accurate sensor data is the foundation for subsequent detection of abnormal signals and comparison of simulation results, untimely or incomplete data acquisition can lead to missed anomalies and deviations in probability updates, affecting the accuracy of early warnings. Therefore, step S1 involves real-time acquisition of the sensor data stream from the fire monitoring area; the specific implementation method is as follows: Three types of sensors are deployed within the target compartments of a nuclear power plant (such as auxiliary compartments in the reactor building): Firstly, distributed fiber optic temperature sensors are laid along the ceiling, walls, and surfaces of key equipment (such as cable trays and pipes) of the compartment, with a sampling interval of 1 meter and a data sampling frequency of 1 Hz, to obtain temperature distribution data of continuous space. Secondly, five point-type photoelectric smoke sensors are deployed at a height of 1.5m at the four corners and the middle of the compartment, with a response threshold of 0.12dB / m and a data upload frequency of 0.5Hz. Thirdly, a gas sensor monitors CO concentration. Its deployment location is consistent with that of the smoke sensor. The measurement range is 0-1000ppm, the accuracy is ±5%, and the data upload frequency is 0.5Hz. All sensors are connected to the data acquisition gateway via industrial Ethernet. The gateway uses an edge computing module to preprocess the raw data (such as removing noise and filling in missing values) and then transmits it to the system's data acquisition module in real time via the MQTT protocol. Through the above steps, multi-parameter, high spatiotemporal resolution monitoring of temperature, smoke, and CO concentration within nuclear power plant compartments is achieved, providing comprehensive and timely data support for subsequent anomaly detection and avoiding early warning delays caused by missing or delayed data.

[0022] Since the appearance of abnormal signals indicates a potential fire, it is necessary to quickly generate initial scenario hypotheses to initiate subsequent simulations. If the scenario hypotheses are singular or unreasonable, it will limit the scope of subsequent extrapolation and increase the risk of missed detections. Using a dynamic probabilistic fire scenario tree, the rationality of the scenario can be quantified through probability, providing a basis for subsequent computing power allocation. Therefore, in step S2, responding to abnormal signals in the sensor data stream, at least one initial fire scenario is generated based on a rapid diagnostic model, serving as the root node of the dynamic probabilistic fire scenario tree. The specific implementation method is as follows: The data acquisition module performs real-time threshold judgment on the received sensor data. When a sensor data exceeds a preset threshold (e.g., a fiber optic temperature sensor detects a sudden temperature rise exceeding 50℃ / min, or a smoke sensor signal reaches the alarm threshold), it is determined to be an abnormal signal. At this time, a rapid diagnostic model is activated. This model uses a machine learning classifier based on random forest. The input parameters include: the type of sensor that triggered the anomaly (e.g., fiber optic temperature sensor), location (e.g., the northwest corner of the compartment ceiling, coordinates X=2m, Y=3m, Z=5m), signal amplitude (e.g., a sudden temperature rise of 60℃ / min), and rate of change (e.g., the temperature change rate increases from 10℃ / min to 60℃ / min within 5 seconds). The model is trained offline (the training dataset contains sensor response data under different fire source locations and power levels within the nuclear power plant compartment), and outputs three initial fire scenario hypotheses as the root node: Scenario 1: The fire source is located at the location of the anomaly sensor (X=2m, Y=3m, Z=5m), the initial fire source power is 500kW, and the fire growth model is the t² rapid growth model (growth coefficient 0.0469kW / s²). Scenario 2: The fire source is located in the area adjacent to the abnormal sensor (X=3m, Y=3m, Z=5m), the initial fire source power is 300kW, and the fire growth model is a t² medium-speed growth model (growth coefficient 0.0117kW / s²). Scenario 3: The fire source is located in the cable tray inside the compartment (X=4m, Y=2m, Z=3m). The initial fire source power is 400kW. The fire growth model is a t² rapid growth model. The initial real-time probability value of each root node is set to 1 / 3 (because there is no more data to distinguish the rationality of the scenario in the initial stage). Through the above steps, multiple initial scenarios covering possible fire situations are quickly generated, avoiding the limitations of a single hypothesis; a machine learning classifier is used to ensure the correlation between scenario hypotheses and sensor anomaly features, improving the rationality of the initial scenarios; the scenario tree structure lays the foundation for subsequent dynamic evolution and computing power scheduling.

[0023] Considering that CFD models can reproduce the physical processes of fire (such as temperature field and smoke field distribution) with high fidelity, they are the core tool for predicting fire situations; parallel computing can process multiple scenarios simultaneously, improving simulation efficiency. If serial computing is used, the simulation progress will lag behind the real time, losing its early warning value. Therefore, step S3, using computational fluid dynamics (CFD) models, performs parallel simulation calculations on the fire scenarios corresponding to each active node in the dynamic probabilistic fire scenario tree; the specific implementation method is as follows: The computation scheduling engine controls a parallel computing cluster (containing 20 computing nodes, each configured with two 16-core CPUs and 128GB of memory), using FDS (Fire Dynamics Simulator) as the CFD model (FDS is a mature tool in the field of fire simulation that supports parallel computing); for the fire scenarios corresponding to the three root nodes, FDS computation models are constructed respectively: First, the computational domain was set to the actual dimensions of a nuclear power plant compartment (8m long × 6m wide × 5m high), and the mesh was divided using a structured mesh with a mesh size of 0.2m × 0.2m × 0.2m (balancing computational accuracy and speed). Second, the boundary conditions are set to the actual wall material of the compartment (e.g., concrete, with a thermal conductivity of 1.7 W / (m²)). K), specific heat capacity 900 J / (kg) K)) and ventilation conditions (e.g., one air supply outlet in the compartment with a wind speed of 1 m / s; one exhaust outlet with an air volume of 500 m³ / h). Third, the fire source settings are configured according to the initial fire source power and growth model of each scenario; The computation scheduling engine allocates the three FDS model tasks to different computing nodes and uses the MPI (Message Passing Interface) parallel communication protocol to realize data interaction between computing nodes. The simulation time step is set to 0.1s. Every 10 time steps (i.e., 1 second of real time), the simulation field data (including temperature field, flue gas concentration field, and thermal radiation flux field data) of a specific time step is output and transmitted to the scene management engine. At the same time, since there is no difference in scheduling priority at present (the probability is the same in the initial stage), the same computing resources are temporarily allocated to each scene according to the principle of average allocation (each scene occupies the computing power of about 3 computing nodes). Through the above steps, physical field evolution data of multiple initial fire scenarios are obtained simultaneously through parallel CFD simulation, providing a basis for subsequent scenario expansion; the high fidelity of the FDS model ensures that the simulation results can reflect the real fire physical process, providing a reliable physical basis for early warning; parallel computing improves the simulation speed and initially avoids simulation delays.

[0024] Because fire evolution is uncertain (e.g., changes in combustible material distribution or ventilation conditions may lead to sudden changes in fire intensity), simulation based solely on the initial scenario cannot cover all possible evolution paths. By deriving sub-scenarios through a set of physical rules, the scenario tree can be dynamically expanded to ensure that no key fire evolution directions are missed. Therefore, in step S4, based on the simulation field data of a specific time step generated by parallel simulation calculation, and based on a preset set of physical rules, sub-scenario nodes of the currently active node are derived and generated to expand the dynamic probabilistic fire scenario tree. The set of physical rules includes one or more preset physical evolution rules. The execution conditions of the physical evolution rules are based on intermediate result parameters generated by the CFD model simulation. The execution action of the physical evolution rules is to generate one or more new sub-scenario nodes. The intermediate result parameters include at least one of predicted temperature, smoke concentration, and thermal radiation flux. The generation of the sub-scenario nodes includes modifying the boundary conditions or fire source items of the fire scenario corresponding to the currently active node, which serves as the starting point for the derivation. The specific implementation method is as follows: The scene management engine has a pre-defined set of physics rules, which includes two core physics evolution rules, specifically: Rule 1: If, at a specific time step in the CFD simulation (a specific time step can be a fixed physical time interval in the CFD simulation, for example, the time point when the simulation progresses by 1 second of real time, or the time point when a significant change in key physical quantities, such as the highest temperature or flue gas layer height, is detected, the system extracts complete simulation field data at this time point, including the component fields of temperature, pressure, concentration, and velocity, as input for scenario deduction; in this embodiment, the specific time step is when the simulation has been running for 10 seconds), the predicted temperature of a certain area exceeds 800℃ (ignition temperature threshold for combustibles (such as cable insulation layers)), and there is a pre-set combustible distribution in this area (pre-entered into the system through nuclear power plant compartment design drawings), then a sub-scenario node is generated, the fire source item of the original scenario is modified, and a secondary fire source (with a power of 30% of the original fire source power) is added to this area. Rule 2: If the simulated flue gas concentration (CO concentration) exceeds 500 ppm in the air outlet area, and the simulation shows that the air outlet wind speed decreases by more than 30% due to flue gas obstruction, a sub-scene node is generated, and the boundary conditions of the original scene are modified to adjust the air outlet wind speed to 70% of the original wind speed. Taking the initial scene 1 as an example, after 10 seconds of simulation, the simulation field data shows that the temperature in the area X=2.5m, Y=3m, Z=5m reaches 850℃, and there is a cable tray (combustible material) in this area, triggering Rule 1 and generating a sub-scene node, namely Scene 1-1, adding a secondary fire source with 30% power (150kW); at the same time, the CO concentration in the air outlet area reaches 520 ppm, and the wind speed drops from 1m / s to 0.65m / s, triggering Rule 2 and generating a sub-scene node, namely Scene 1-2, adjusting the air outlet wind speed to 0.7m / s. Similarly, scenarios 2 and 3 in the initial scenario are evaluated. Scenario 2 is not generated as the simulated temperature does not reach the threshold and the wind speed change is small. Scenario 3 simulation shows that the temperature in the area with X=4.5m, Y=2m, and Z=3m reaches 820℃, indicating the presence of combustible materials, so a sub-scenario node, namely Scenario 3-1, is generated. At this point, the active nodes of the dynamic probabilistic fire scenario tree include 3 root nodes and 3 sub-scenario nodes (Scenario 1-1, Scenario 1-2, and Scenario 3-1). Through the above steps, the scenario tree is dynamically expanded based on the physical laws of fire, covering key evolution paths such as fire spread and ventilation changes, avoiding early warning omissions due to fixed scenarios; the generation of sub-scenarios is based on specific simulation data, ensuring the rationality of scenario evolution and providing support for subsequent accurate early warning.

[0025] Considering that real-time sensor data can reflect the actual characteristics of a fire as it develops, the degree of consistency between the scenario and the real fire can be determined by comparing the data with simulated data from various scenarios. The probability value is updated based on this consistency to quantify the rationality of each scenario, providing a core basis for subsequent pruning and computational power scheduling. Failure to update the probability will lead to a deviation of computational power allocation from actual needs. Therefore, in step S5, the predicted data simulated by the CFD model corresponding to each active node is periodically compared with the real-time sensor data, and the real-time probability value of each active node is dynamically updated based on the comparison results. The real-time probability value of each active node is updated using a Bayesian update method. Specifically, this involves: calculating the likelihood of each active node based on the degree of agreement between the predicted data simulated by the CFD model and the actual measured data from the corresponding location sensor; where a higher degree of agreement results in a higher likelihood; multiplying the current real-time probability value of each active node by its corresponding likelihood to obtain the updated probability weight; and normalizing the updated probability weights of all active nodes so that the sum of the probability weights of all active nodes is 1, thus obtaining the new real-time probability value for each node. The specific implementation is as follows: Set the data comparison period to 5 seconds (i.e., execute step S5 once every 5 seconds); The scene management engine obtains the latest real-time sensor data within the last 5 seconds from the data acquisition module (such as the average temperature of the fiber optic temperature sensor at each location, the signal value of the smoke sensor, and the concentration value of the CO sensor), and at the same time extracts the prediction data of the CFD simulation of each active node at the corresponding time step (such as the temperature, smoke concentration, and CO concentration of the same sensor location in the simulation). The real-time probability values ​​are updated using a Bayesian update-based method: First, calculate the likelihood. Taking temperature data as an example, if the error between the simulated predicted temperature and the actual measured temperature of a certain node is within ±5%, the likelihood is set to 1.2; if the error is between 5% and 10%, the likelihood is set to 1.0; if the error exceeds 10%, the likelihood is set to 0.8. Similarly, calculate the likelihood for smoke concentration (with an error of ±10%) and CO concentration (with an error of ±8%), and take the average of the three as the final likelihood of the node. Taking node scenario 1 as an example, its temperature error is 3%, smoke concentration error is 8%, CO concentration error is 5%, and the average likelihood is 1.1; for node scenario 1-1, the temperature error is 6%, smoke concentration error is 9%, CO concentration error is 7%, and the average likelihood is 1.0; for node scenario 1-2, the temperature error is 4%, smoke concentration error is 11%, CO concentration error is 6%, and the average likelihood is 0.95; for node scenario 2, the temperature error is 12%, smoke concentration error is 15%, CO concentration error is 11%, and the average likelihood is 0.75; for node scenario 3, the temperature error is 5%, smoke concentration error is 7%, CO concentration error is 4%, and the average likelihood is 1.15; for node scenario 3-1, the temperature error is 7%, smoke concentration error is 8%, CO concentration error is 6%, and the average likelihood is 1.05.

[0026] Secondly, calculate the updated probability weights. The current real-time probability value of each node is multiplied by its own likelihood. Initially, the probability of each root node is 1 / 3. Since child nodes are newly generated, their initial probabilities are distributed as 1 / 2 of the parent node's probability (e.g., the initial probabilities of nodes Scenario 1-1 and Scenario 1-2 are both (1 / 3) × 1 / 2 = 1 / 6). Therefore, the weight of node Scenario 1 is: (1 / 3) × 1.1 ≈ 0.3667; the weight of node Scenario 1-1 is: (1 / 6) × 1.0 ≈ 0.1667; the weight of node Scenario 1-2 is: (…). 1 / 6) × 0.95 ≈ 0.1583; Node Scenario 2 weight: (1 / 3) × 0.75 = 0.25; Node Scenario 3 weight: (1 / 3) × 1.15 ≈ 0.3833; Node Scenario 3-1 weight: (1 / 6) × 1.05 ≈ 0.175. It should be further noted that the above initial probability allocation method is only an example. In actual applications, the rapid diagnostic model can also output the initial confidence level of each scenario hypothesis as its probability, and the child node probability can also adopt other inheritance strategies.

[0027] Finally, after normalization, the sum of the weights of all nodes is... The new real-time probability value of each node is its own weight divided by 1.5. Finally, the probability values ​​are: Node Scenario 1 ≈ 0.244, Node Scenario 1-1 ≈ 0.111, Node Scenario 1-2 ≈ 0.105, Node Scenario 2 ≈ 0.167, Node Scenario 3 ≈ 0.255, and Node Scenario 3-1 ≈ 0.117.

[0028] Through the above steps, Bayesian updates are used to dynamically correct the probabilities of each scenario based on real-time data, so that the probability values ​​can accurately reflect the degree of consistency between the scenario and the actual fire situation. The quantified probability values ​​provide an objective basis for subsequent pruning (eliminating low-probability nodes) and computing power scheduling (prioritizing high-probability nodes), thereby improving the system's adaptability to dynamic fire situations.

[0029] Considering that low-probability nodes or nodes with large deviations between simulated and measured data are highly unlikely to reflect the true fire situation, continuing to allocate computing power to them would waste resources and may interfere with early warning judgments. Timely pruning and termination of calculations can reclaim computing power for high-value nodes, improve computing power utilization efficiency, and ensure that early warning resources are focused on key scenarios. Therefore, in step S6, according to preset elimination conditions, active nodes and their descendant nodes with real-time probability values ​​below the first threshold or deviations between predicted data and sensor data continuously exceeding the second threshold are pruned, and their corresponding CFD model simulation calculation tasks are terminated. The elimination conditions include: Condition 1, the real-time probability value of the node is below the first threshold for N consecutive update cycles; Condition 2, the error between the key parameters predicted by the node's CFD model simulation and the corresponding sensor measured data exceeds the second threshold for M consecutive update cycles. Pruning is triggered when either Condition 1 or Condition 2 is met, where N and M are preset positive integers. The specific implementation method is as follows: Preset elimination conditions: The first threshold is set to 0.1 (real-time probability value threshold), the second threshold is set to 15% (deviation threshold between predicted data and sensor data), N=2 (for two consecutive update cycles), and M=2 (for two consecutive update cycles). The scene management engine judges each active node: Node 1-2 has a current real-time probability value of 0.105 (close to the first threshold), and in the previous two update cycles (a total of 10 seconds), the deviation between its simulated data and measured data is 14% and 16% respectively (one cycle exceeds the second threshold), so it does not meet the pruning conditions; Node 2 has a current real-time probability value of 0.167 (higher than the first threshold), but the deviation in the previous two update cycles is 18% and 20% respectively (two consecutive cycles exceed the second threshold), meeting condition two, and triggering the pruning operation; no node meets condition one (probability below 0.1 for two consecutive cycles).

[0030] The pruning operation is as follows: the scene management engine sends a pruning instruction to the computing scheduling engine, specifying node scene 2; the computing scheduling engine immediately terminates the FDS simulation computing task corresponding to node scene 2, releasing the computing power of the 3 computing nodes it occupies; at the same time, it deletes node scene 2 and its possible descendant nodes (in this embodiment, node scene 2 has no descendant nodes) from the dynamic probabilistic fire scene tree, and will not perform simulation, update or other operations on it in the future. Through the above steps, nodes that deviate significantly from the actual fire situation are eliminated in a timely manner, avoiding waste of computing resources; the recovered computing power can be redistributed to provide more resource support for high-priority nodes, improving the overall simulation efficiency and early warning accuracy; and reducing the number of invalid nodes in the scene tree, thereby reducing the complexity of system data processing and management.

[0031] Considering that different active nodes have different values ​​for early warning (high-probability nodes are more likely to reflect the real fire situation, and lagging simulation nodes will lose their timeliness if they are not accelerated), priority scheduling can tilt computing power towards high-value nodes; recovering and redistributing the computing power of pruned nodes can maximize computing power utilization efficiency, avoid resource idleness, and ensure that the simulation progress of key scenarios keeps up with real time. Therefore, step S7 determines the scheduling priority of each active node based on the real-time probability value of each active node and the degree to which the corresponding CFD model simulation progress lags behind real time, and dynamically allocates computing resources to active nodes in the dynamic probability fire scenario tree according to the scheduling priority; wherein, the computing resources recovered from pruned nodes are redistributed to active nodes with higher scheduling priority; the scheduling priority of each active node is determined based on its real-time probability value and the degree to which its CFD model simulation progress lags behind real time; wherein, the higher the real-time probability value of a node and the greater the degree to which its CFD model simulation progress lags behind real time, the higher its scheduling priority; the determination of the scheduling of each active node Priority is also based on a critical flag: when the fire scenario represented by a node involves an impact on predefined safety-critical equipment or areas, the critical flag of that node is activated; nodes with activated critical flags have their scheduling priority increased under the same conditions; the specific steps for dynamically allocating computing resources include: S71, obtaining the scheduling priorities of all currently active nodes and sorting them according to priority; S72, dividing the available CFD model computing core resources into multiple computing resource units according to a preset ratio; S73, allocating the first-ranked active node with a first proportion of computing resource units, and allocating the remaining computing resource units according to the scheduling priority of each node; S74, periodically or upon receiving a node pruning notification, re-executing steps S71 to S73; for different active nodes, dynamically adjusting the computational precision of their CFD models according to their scheduling priorities; wherein, nodes with scheduling priorities higher than the fourth threshold use the first computational precision, and nodes with scheduling priorities lower than the fourth threshold use the second computational precision, which is lower than the first computational precision; the specific implementation method is as follows: First, determine the scheduling priority. The priority is calculated using the following formula: ,in, This is the real-time probability value. The maximum real-time probability value. for , for , The critical correction value is the difference between the actual time and the simulated completion time (e.g., if the current actual time is 30s, and a certain node is simulated to be completed in 25s, the lag time is 5s). The maximum lag time is set to 10s (if it exceeds this, the warning becomes meaningless). The critical correction value includes: if the node scenario involves safety-critical equipment in a nuclear power plant compartment (e.g., emergency cooling pipes, whose location and importance are predefined in the system), the correction value is 0.1; otherwise, it is 0. The current real-time probability value of each active node (Scenario 1, Scenario 1-1, Scenario 1-2, Scenario 3, Scenario 3-1) is the highest at 0.255 for Scenario 3. The simulation progress lag times are as follows: Scenario 1 lags by 4 seconds, Scenario 1-1 by 5 seconds, Scenario 1-2 by 6 seconds, Scenario 3 by 3 seconds, and Scenario 3-1 by 4 seconds. Among them, Scenario 3 and Scenario 3-1 involve emergency cooling pipelines (critical flag activated), with a correction value of 0.1, while the correction value for the other nodes is 0. It should be further noted that the system maintains a predefined database of safety-critical equipment and areas, which stores equipment identifiers, spatial coordinates, and criticality levels. When the predicted impact range of a node scenario spatially overlaps with an entry in this database, it is determined to involve critical equipment, and the critical flag of that node is activated. Calculate the priority of each node: Node Scenario 1: ; Node Scenario 1-1: ; Node Scenario 1-2: ; Node Scenario 3: ; Node Scenario 3-1: The priority order is: Node Scenario 3 (0.98) > Node Scenario 1 (0.816) > Node Scenario 3-1 (0.614) > Node Scenario 1-1 (0.462) > Node Scenario 1-2 (0.407). Next, dynamic computing power allocation is performed. The total number of computing cores in the parallel computing cluster is 20 × 16 = 320, which is divided into 32 computing resource units (10 cores per unit). The 3 units released by node scenario 2 (originally occupying 30 cores) are reclaimed, and there are currently 32 available units (with no other occupied cores). It should be further noted that the dynamic computing power allocation strategy can be flexibly set according to the actual architecture of the computing cluster. For example, it can be divided by the number of physical cores, the number of virtual CPUs, or the entire computing node. The above example of 10 cores per unit is just an example. Distribute according to a preset ratio: The first ratio is set to 40% (the first-ranked node occupies 40% of the available units), and node scenario 3 is allocated 32 × 40% ≈ 13 units (130 cores). The remaining units are 32-13=19, allocated according to the priority ratio of each node: Node scenario 1 priority ratio is 0.816 / (0.816+0.614+0.462+0.407)≈0.816 / 2.299≈0.355, allocated 19×0.355≈7 units (70 cores); Node scenario 3-1 priority ratio is 0.614 / 2.299≈0.267, allocated 19×0.267≈5 units (50 cores); Node scenario 1-1 priority ratio is 0.462 / 2.299≈0.201, allocated 19×0.201≈4 units (40 cores); Node scenario 1-2 priority ratio is 0.407 / 2.299≈0.177, allocated 19×0.177≈3 units (30 cores). Meanwhile, a fourth threshold is set to 0.7 (scheduling priority threshold). Nodes with a priority higher than 0.7 (node ​​scenario 3, scenario 1) use the first calculation precision (FDS model mesh size 0.15m×0.15m×0.15m, time step 0.05s), while nodes with a priority lower than 0.7 (node ​​scenario 3-1, scenario 1-1, scenario 1-2) use the second calculation precision (mesh size 0.25m×0.25m×0.25m, time step 0.15s). Subsequently, priority calculation and computing power allocation are re-executed every 5s (synchronized with the probability update cycle) or when a pruning notification is received.

[0032] Through the above steps, a composite priority strategy is used to ensure that high-probability, critical scenarios, and nodes with simulation lags receive more computing power, guaranteeing their simulation progress and accuracy, and meeting the requirements for timely and accurate early warning. Differentiated computing precision achieves a balance between speed and accuracy under computing power constraints. High-value nodes use high precision to ensure accurate predictions, while low-value nodes use low precision to save computing power. The dynamic allocation and recycling mechanism of computing power maximizes resource utilization efficiency and avoids idleness and waste.

[0033] Since fire is a dynamic evolutionary process, it is necessary to continuously track fire changes through a cycle of simulation, updating, pruning, and scheduling to ensure that the scene tree always reflects the latest fire probability. Outputting early warnings based on the simulation results of high-probability nodes can provide specific and reliable basis for emergency decision-making. Therefore, in step S8, steps S3 to S7 are repeated, and fire development prediction and early warning information are generated and output based on the simulation results of the fluid dynamics model corresponding to active nodes whose current real-time probability value is higher than the third threshold. The specific implementation method is as follows: The system repeats steps S3 to S7 in 5-second cycles. In subsequent iterations, nodes 1-2 are pruned because their probability continuously drops to 0.09 (below the first threshold for two consecutive cycles), and their computing power is reclaimed and allocated to node 3. The simulation progress of node 3 gradually catches up with the real time. The simulation results show that within the next 10 seconds, the temperature in the emergency cooling pipe area will rise to 600℃ (exceeding the pipe's tolerance temperature of 500℃), and the CO concentration will diffuse into the personnel evacuation passage (reaching a concentration of 800ppm). The third threshold is set to 0.2 (the threshold for high-probability nodes). The current real-time probability value of node 3 rises to 0.32 (above the third threshold), while the probability of node 1 drops to 0.18 (below the threshold). The early warning output module generates fire development prediction and early warning information based on the CFD simulation results of node scenario 3, including: First, fire development prediction: within the next 10 seconds, the fire source (X=4m, Y=2m, Z=3m) will spread to the emergency cooling pipe (X=5m, Y=2m, Z=3m), the temperature in the pipe area will reach 600℃, and the CO concentration will spread to the evacuation route on the east side. Second, the warning level: Level II warning (set according to the warning level of nuclear power plants, with Level I being the most severe). Third, emergency recommendations: Immediately activate the emergency cooling pipeline system, close the evacuation route on the east side, and guide personnel to evacuate through the west side route; Early warning information is simultaneously output through industrial control screens, audible and visual alarms, and emergency command terminals.

[0034] Through the above steps and continuous iteration, the system can dynamically track changes in the fire situation and adjust the scene tree and computing power allocation in a timely manner. Based on the simulation results of high-probability nodes, it outputs specific fire predictions and emergency suggestions, providing accurate decision support for on-site personnel and effectively improving the fire emergency response capabilities of high-risk industrial facilities.

[0035] Example 2 This embodiment describes the application of the fire early warning method based on computing power allocation in fire early warning of large chemical storage tank areas to illustrate the universality of the method. The specific implementation steps are as follows: Considering the large area and numerous tanks in large chemical storage tank areas (such as gasoline storage tank areas), and the rapid spread of fires, large-scale, multi-parameter real-time monitoring is required to promptly detect initial fires and provide a data foundation for subsequent early warnings. Therefore, step S1, real-time acquisition of sensor data streams, is specifically implemented as follows: Sensors were deployed in the tank area (containing 10 gasoline tanks of 5000m³ each, arranged in 2 rows and 5 columns, with a tank spacing of 10m): Firstly, four infrared thermal imaging cameras are deployed at high points around the tank area, covering all tanks with a sampling frequency of 0.2Hz, to acquire temperature distribution image data of the tank surface and the area between tanks. Secondly, point-type flame detectors are installed on the top of each storage tank and in the middle of the passage between tanks, totaling 30. The response wavelength is 2.5-4.0μm, the alarm delay is ≤0.5s, and the data upload frequency is 1Hz. Thirdly, wind speed and direction sensors are installed in five locations at the center and four corners of the storage tank area. The measurement range is 0-30m / s (wind speed) and 0-360° (wind direction), with an accuracy of ±0.3m / s and ±3°, and a data upload frequency of 1Hz. All sensors are connected to the data acquisition gateway via a wireless LoRa network (to meet the needs of large-scale, low-power storage tank areas). The gateway converts the data format (e.g., converting thermal imaging image data into temperature matrix data) and then transmits it to the system data acquisition module via a 4G network. Through the above steps, large-scale monitoring of temperature, flame, wind speed and direction in large chemical storage tank areas is achieved. Infrared thermal imaging can quickly detect local high temperatures on the surface of the tank (characteristics of the early stage of a fire), and wind speed and direction data provide support for the wind field boundary conditions in subsequent CFD simulations, thus meeting the monitoring needs for fire early warning in storage tank areas.

[0036] Since the initial stage of a tank area fire may manifest as localized overheating or small flames on the tank surface, after an abnormal signal appears, it is necessary to generate an initial scenario based on the tank layout and medium characteristics (gasoline is flammable and volatile) to ensure coverage of possible ignition tanks and the scale of the fire source, laying the foundation for subsequent simulations. Therefore, step S2: Responding to the abnormal signal to generate an initial fire scenario, specifically implemented as follows: When an infrared thermal imaging camera detects a localized temperature of 300℃ on the wall of a storage tank (such as tank No. 3) (the auto-ignition temperature of gasoline is approximately 280℃; exceeding this temperature is considered abnormal), or a flame detector triggers an alarm, a rapid diagnostic model is activated. Unlike Implementation 1, this model employs a rule-based expert system. Input parameters include: the type of abnormal sensor (such as an infrared thermal imaging camera), its location (north side of tank No. 3, 2m high), signal amplitude (temperature 320℃), and rate of change (temperature rising from 80℃ to 320℃ within 10 seconds). The expert system has a built-in rule base (constructed based on chemical storage tank fire accident cases and theoretical analysis), and outputs two initial fire scenarios as root nodes: Scenario 1: The fire source is located on the north side of the tank wall of storage tank No. 3 (X=30m, Y=20m, Z=2m). The initial fire source power is 1000kW (corresponding to a small jet fire). The fire growth model is an instantaneous growth model (gasoline flames spread quickly and reach a stable power initially). Scenario 2: The fire source is located on the ground between storage tanks 3 and 4 (X=35m, Y=20m, Z=0.5m), with an initial fire power of 800kW (corresponding to flowing fire on the ground). The fire growth model is a linear growth model (growth rate 200kW / min). The initial real-time probability value of each root node is set to 0.5. Through the above steps, based on the characteristics of chemical storage tank areas and expert rules, targeted initial fire scenarios are generated, covering typical tank fire types such as jet fire and flowing fire; the initial probability is evenly distributed, leaving room for subsequent probability correction based on real-time data.

[0037] Subsequently, the core difference between steps S3 to S8 in this embodiment and the process in embodiment 1 lies in: Step S3 (CFD Simulation): FLUENT is used as the CFD model (suitable for large-scale flow field simulation). The computational domain is set as the tank area and a surrounding 50m range (100m long × 80m wide × 30m high). Unstructured meshing is used (the mesh in the tank area is refined to 0.5m, and the mesh in the surrounding area is 1m). Real-time wind speed and direction data are introduced into the boundary conditions (e.g., current wind speed 3m / s, northerly wind, set as the boundary condition at the inlet of the computational domain). The parallel computing cluster is configured with 30 computing nodes, initially allocating 15 nodes of computing power to each scenario. The simulation time step is 0.2s, and the simulation field data is output every 2s. Step S5 (Probability Update): The likelihood calculation adds a wind speed and direction matching index (if the error between the simulated wind field and the measured wind speed and direction is within ±10%, the likelihood is increased by 0.1); due to the large number of sensors in the tank area, a weighted average is used to calculate the likelihood (the weights for temperature, flame signal, and wind speed and direction are 0.4, 0.3, and 0.3, respectively). Step S7 (Computing Power Allocation): The activation condition for the critical flag is set to the fire monitor coverage area of ​​the tank area in the scenario (predefined), and the correction value after activation is 0.15; the first ratio is set to 35%, the fourth threshold is set to 0.65, and high-priority nodes (such as scenarios involving fire monitor areas) adopt the first calculation accuracy with a grid size of 0.3m. Step S8 (Early Warning Output): Based on the simulation results of high-probability nodes (such as the No. 3 storage tank jet fire scenario with a probability of 0.28), it is predicted that the flame will spread to the tank wall of No. 4 storage tank within the next 15 seconds, with a heat radiation flux of 15kW / m² (exceeding the tank steel plate tolerance limit of 10kW / m²). A Level 1 early warning is output, and it is recommended to immediately activate the fire monitors of No. 3 and No. 4 storage tanks and evacuate personnel around the tank area. Through the above steps, the CFD model, probability calculation indicators, and computing power allocation parameters are adjusted according to the characteristics of large chemical storage tank areas to ensure that the early warning scheme is adapted to the evolution of fires in storage tank areas and to provide accurate support for fire emergency response in the chemical industry.

[0038] Example 3 Considering that machine learning-based classifiers require extensive training to ensure the reasonableness of the output scenario, insufficient training data or inappropriate feature selection can lead to biased scenario assumptions. Optimizing the training process and feature input can improve the model's adaptability to different fire situations and ensure the accuracy of the initial scenario assumptions. Therefore, in step S2, the rapid diagnostic model is either a machine learning-based classifier or a rule-based expert system. Its input includes at least the sensor type, location, signal amplitude, and rate of change of the sensor that triggered the abnormal signal, and its output is at least one fire scenario assumption, which includes the fire source location, initial fire source power, and fire growth model parameters. The specific implementation method is as follows: Taking the nuclear power plant compartment fire early warning scenario of Example 1 as an example, the rapid diagnostic model uses an improved random forest classifier, and the specific steps are as follows: First, the training dataset was constructed: fire test data under different operating conditions in nuclear power plant compartments were collected (a total of 1000 sets). Each set of data included sensor features (type, location, signal amplitude, rate of change) and corresponding real fire scene parameters (fire source location, power, growth model); noise data (such as sensor false alarm signals) was also added to improve the model's anti-interference ability. Secondly, feature engineering: In addition to the basic input features, new "sensor combination response features" (such as the correlation between the signal changes of the temperature sensor and the CO sensor) and "environmental parameter features" (such as the initial temperature and humidity of the compartment) are added, for a total of 6 input features. Principal component analysis (PCA) is used to reduce the dimensionality to 4 core features, thereby reducing the computational load of the model. Then, model training: the model is trained using 5-fold cross-validation, and the number of decision trees in the random forest is adjusted (set to 200 trees) and the maximum tree depth is set to 10 to avoid overfitting; Finally, model deployment: The trained model is deployed to the edge computing module of the scene management engine, and TensorRT is used to accelerate the model to ensure that after an abnormal signal is triggered, the model can output three initial fire scenario hypotheses within a short time (e.g., 1 second). Through the above steps, by optimizing training data and feature engineering, the scene prediction accuracy and anti-interference ability of the machine learning classifier are improved; the accelerated model processing ensures the rapid generation of the initial scene, meeting the real-time early warning requirements for response speed.

[0039] Example 4 Considering that heat radiation flux is a key factor in the spread of fire to adjacent combustibles, existing rules only consider temperature and smoke concentration, failing to cover the impact of heat radiation and potentially overlooking secondary fire scenarios caused by heat radiation. Adding new heat radiation-related rules can more comprehensively cover the fire evolution path and improve the completeness of the scenario tree. The specific implementation method is as follows: Add rule 3 to the pre-set physics rule set: If, at a specific time step in the CFD simulation, the predicted thermal radiation flux of a certain area exceeds 20 kW / m² for 5 seconds (the ignition thermal radiation flux threshold for common combustibles such as wood and plastic), and there are pre-set combustibles (such as cables and insulation materials) in that area, then generate a sub-scene node, modify the fire source item of the original scene, and add a secondary fire source (with a power of 25% of the original fire source power) in that area. Taking scenario 3-1 (the jet fire scenario of tank No. 3) of the large chemical storage tank area fire early warning in Example 2 as an example, after 20 seconds of simulation, the simulation field data shows that the heat radiation flux of the south wall of tank No. 4 reaches 22kW / m² and lasts for 5 seconds. Since there is tank insulation material (combustible material) in this area, rule 3 is triggered, generating sub-scenario node 3-1-1. A secondary fire source of 250kW (25% of the original fire source power of 1000kW) is added to the south wall of tank No. 4 (X=36m, Y=20m, Z=2m). Simultaneously, the boundary conditions of this sub-scenario (such as wind speed and direction remaining northerly at 3m / s) are recorded to ensure consistency in subsequent simulations. Through the above steps, the evolution path of secondary fires caused by thermal radiation is covered by the newly added thermal radiation triggering rules, avoiding the omission of key fire development directions in the scene tree; the sub-scene generation is based on specific thermal radiation data and combustible material distribution, ensuring the physical rationality of scene evolution and providing support for accurate prediction of the fire spread range.

[0040] Example 5 To implement the methods of Embodiments 1 to 4, please refer to... Figure 2 As shown, the purpose of Embodiment 2 is to provide a fire early warning system based on computing power allocation. This system is deployed on a high-performance edge server and includes: The data acquisition module is configured to acquire sensor data streams from the fire monitoring area in real time. The data acquisition module consists of an MQTT message queue service program, which listens on the port to receive sensor JSON data packets from the field PLC. The scene management engine, connected to the data acquisition module, is used to respond to abnormal signals in the sensor data stream and generate at least one initial fire scene based on a rapid diagnostic model, serving as the root node of the dynamic probabilistic fire scene tree. It periodically compares the predicted data simulated by the CFD model corresponding to each active node with the real-time acquired sensor data, and dynamically updates the real-time probability value of each active node based on the comparison results. According to preset elimination conditions, it prunes active nodes and their descendant nodes whose real-time probability values ​​are below a first threshold or whose deviation between predicted data and sensor data consistently exceeds a second threshold. The scene management engine is a daemon running in the background, internally maintaining a tree data structure containing a Bayesian update algorithm module and a pruning logic judgment module. Furthermore, to prevent single-point failures, the scene management engine supports hot-swapping between primary and backup engines; when the primary engine fails, the backup engine can restore the scene tree state from the Redis database and continue running. A computational scheduling engine, connected to the scene management engine, controls the parallel computing cluster to perform parallel simulation calculations of the fire scenarios corresponding to each active node in the dynamic probabilistic fire scenario tree using a CFD model, and feeds back the simulation field data at a specific time step to the scene management engine; based on the simulation field data and a preset set of physical rules, it infers and generates sub-scene nodes of the currently active nodes to expand the dynamic probabilistic fire scenario tree; based on the real-time probability value of each active node obtained from the scene management engine and the degree to which the CFD model simulation progress lags behind the actual time, it determines the scheduling priority of each active node; and dynamically schedules the fire scenarios according to the scheduling priority. The computing resources of the parallel computing cluster are allocated to active nodes in the dynamic probabilistic fire scenario tree. In response to the pruning instructions of the scenario management engine, the computing tasks of the corresponding nodes are terminated and the released resources are reallocated. The computing scheduling engine is a cluster management middleware based on Slurm or Kubernetes. It is responsible for encapsulating CFD solvers (such as OpenFOAM) in Docker containers. When it receives the "expand" instruction from the scenario management engine, it starts a new container. When it receives the "terminate" instruction, it kills the corresponding container ID and dynamically adjusts the released CPU quota to high-priority containers through cgroups. The early warning output module is connected to the scene management engine. Based on the CFD model simulation results of active nodes whose current real-time probability value is higher than the third threshold, it generates and outputs fire development prediction and early warning information. The early warning output module is a Websocket service that pushes JSON format early warning messages to the monitoring screen and drives the relay interface of the on-site audible and visual alarm.

Claims

1. A fire early warning method based on computing power allocation, characterized in that, Includes the following steps: S1. Real-time acquisition of sensor data streams in the fire monitoring area; S2. In response to abnormal signals in the sensor data stream, generate at least one initial fire scenario based on the rapid diagnostic model, which serves as the root node of the dynamic probabilistic fire scenario tree. S3. Using a computational fluid dynamics model, perform parallel simulation calculations on the fire scenarios corresponding to each active node in the dynamic probabilistic fire scenario tree; S4. Based on the simulated field data at a specific time step generated by parallel simulation calculation, and based on a preset set of physical rules, deduce and generate sub-scene nodes of the currently active nodes to expand the dynamic probability fire scene tree. S5. Periodically compare the predicted data simulated by the computational fluid dynamics model corresponding to each active node with the real-time sensor data, and dynamically update the real-time probability value of each active node based on the comparison results. S6. Based on the preset elimination conditions, prune active nodes and their descendant nodes whose real-time probability values ​​are lower than the first threshold or whose deviation between predicted data and sensor data is consistently higher than the second threshold, and terminate the computational fluid dynamics model simulation task corresponding to them. S7. Based on the real-time probability value of each active node and the degree to which the simulation progress of the corresponding computational fluid dynamics model lags behind the real time, determine the scheduling priority of each active node, and dynamically allocate computing resources to the active nodes in the dynamic probability fire scenario tree according to the scheduling priority; wherein, the computing resources reclaimed from the pruned nodes are redistributed to active nodes with higher scheduling priority. S8. Repeat steps S3 to S7, and generate and output fire development prediction and early warning information based on the simulation results of the fluid dynamics model corresponding to the active nodes whose current real-time probability value is higher than the third threshold.

2. The fire early warning method based on computing power allocation according to claim 1, characterized in that, In step S2, the rapid diagnostic model is a machine learning-based classifier or a rule-based expert system; its input includes at least the sensor type, location, signal amplitude, and rate of change of the sensor that triggers the abnormal signal, and its output is at least one fire scenario hypothesis, which includes the fire source location, initial fire source power, and fire growth model parameters.

3. The fire early warning method based on computing power allocation according to claim 1, characterized in that: In step S4, the physical rule set includes one or more preset physical evolution rules; The execution conditions of the physical evolution rule are based on intermediate result parameters generated by the computational fluid dynamics model simulation. The execution action of the physical evolution rule is to generate one or more new sub-scene nodes. The intermediate result parameters include at least one of predicted temperature, smoke concentration, and thermal radiation flux. The generation of the sub-scene node includes modifying the boundary conditions or fire source terms of the fire scene corresponding to the currently active node, which serves as the starting point for the deduction.

4. The fire early warning method based on computing power allocation according to claim 1, characterized in that, In step S5, the real-time probability value of each active node is dynamically updated based on the comparison results. This is achieved using a Bayesian update-based method, specifically including: The likelihood of each active node is calculated based on the degree of agreement between the predicted data from the computational fluid dynamics model simulation and the measured data from the corresponding position sensor; the higher the degree of agreement, the higher the likelihood. Multiply the current real-time probability value of each active node by its corresponding likelihood to obtain the updated probability weight; The updated probability weights of all active nodes are normalized so that the sum of the probability weights of all active nodes is 1, thus obtaining the new real-time probability value of each node.

5. The fire early warning method based on computing power allocation according to claim 1, characterized in that, In step S6, the elimination conditions include: Condition 1: The real-time probability value of the node is lower than the first threshold for N consecutive update cycles; Condition 2: The error between the key parameters simulated and predicted by the nodal computational fluid dynamics model and the corresponding measured data from the sensors exceeds the second threshold for M consecutive update cycles; The pruning operation is triggered when either condition one or condition two is met, where N and M are preset positive integers.

6. The fire early warning method based on computing power allocation according to claim 1, characterized in that, In step S7, the scheduling priority of each active node is determined based on its real-time probability value and the degree to which the progress simulated by the computational fluid dynamics model lags behind the actual time. The higher the real-time probability value of a node, and the greater the lag between its computational fluid dynamics model simulation progress and real time, the higher its scheduling priority.

7. The fire early warning method based on computing power allocation according to claim 6, characterized in that, The determination of the scheduling priority of each active node is also based on key flag bits: When the fire scenario represented by a node involves an impact on predefined safety-critical equipment or areas, the critical flag of that node is activated. Nodes whose critical flags are activated will have their scheduling priority increased under the same conditions.

8. The fire early warning method based on computing power allocation according to claim 1, characterized in that, In step S7, the specific steps for dynamically allocating computing resources include: S71. Obtain the scheduling priority of all currently active nodes and sort them according to priority. S72. Divide the available computational fluid dynamics model computational core resources into multiple computational resource units according to a preset ratio; S73. Allocate the first-ranked active node with the first proportion of computing resource units, and allocate the remaining computing resource units proportionally according to the scheduling priority of each node. S74. Periodically or upon receiving a node pruning notification, re-execute steps S71 to S73.

9. The fire early warning method based on computing power allocation according to claim 1, characterized in that, In step S3, the computational accuracy of the computational fluid dynamics model of different active nodes is dynamically adjusted according to their scheduling priority; wherein, nodes with scheduling priority higher than the fourth threshold adopt the first computational accuracy, and nodes with scheduling priority lower than the fourth threshold adopt the second computational accuracy, which is lower than the first computational accuracy.

10. A fire early warning system based on computing power allocation, wherein the fire early warning system based on computing power allocation is used to implement the fire early warning method based on computing power allocation according to any one of claims 1 to 9, characterized in that, include: The data acquisition module is configured to acquire sensor data streams from the fire monitoring area in real time. The scene management engine, connected to the data acquisition module, is used to respond to abnormal signals in the sensor data stream, generate at least one initial fire scene based on a rapid diagnostic model, and serve as the root node of the dynamic probabilistic fire scene tree; periodically compare the predicted data simulated by the computational fluid dynamics model corresponding to each active node with the real-time acquired sensor data, and dynamically update the real-time probability value of each active node based on the comparison results; according to preset elimination conditions, prune active nodes and their descendant nodes whose real-time probability values ​​are lower than a first threshold or whose deviation between the predicted data and the sensor data is consistently higher than a second threshold. The computation scheduling engine, connected to the scene management engine, is used to control the parallel computing cluster to perform parallel simulation calculations on the fire scenes corresponding to each active node in the dynamic probabilistic fire scene tree using the computational fluid dynamics model, and to feed back the simulation field data at a specific time step to the scene management engine. Based on the simulated field data and the preset physical rule set, the sub-scene nodes of the currently active node are deduced and generated to expand the dynamic probability fire scene tree; Based on the real-time probability value of each active node obtained from the scene management engine and the degree to which the computational fluid dynamics model simulation progress lags behind the real time, the scheduling priority of each active node is determined; according to the scheduling priority, the computing resources of the parallel computing cluster are dynamically allocated to the active nodes in the dynamic probabilistic fire scene tree, and in response to the pruning command of the scene management engine, the computing tasks of the corresponding nodes are terminated and the released resources are reallocated. The early warning output module is connected to the scene management engine. Based on the computational fluid dynamics model simulation results of active nodes whose current real-time probability value is higher than the third threshold, it generates and outputs fire development prediction and early warning information.

Images