Forest fire detection and disappearance method and system

Abstract

The invention discloses a forest fire detection and disappearance method and system, and relates to the technical field of forest fire prevention and data processing, and the method comprises the steps: carrying out the fire inspection of a planned forest region through a plurality of inspection unmanned planes, and transmitting an image and airborne sensor data back to a command center in real time; identifying a fire behavior by using an artificial intelligence method, and performing fire behavior judgment and fire hazard grade division; the command center screens fire points according to the thermal images and updates the positions of the fire points in real time; the fire scene spreading trend is judged according to the returned data, and a basis is provided for the dispatching task of throwing the fire extinguishing bombs by the fire extinguishing unmanned aerial vehicle; using an MAPPO multi-agent reinforcement learning algorithm to train an optimal dynamic fire point collaborative disappearance regulation strategy; and after the open fire is extinguished, the inspection unmanned aerial vehicle continues to inspect to identify a re-ignition point or a smoke point, and secondary large-area fire is prevented. According to the method, efficient collaborative fire extinguishing, dynamic environment adaptation and resource optimization configuration can be achieved, and robustness is kept when part of intelligent agents fail.

Description

Technical Field

[0001] The present invention relates to the field of forest fire detection and extinguishing, and in particular to a forest fire detection and extinguishing method and system. Background Art

[0002] Forests are important ecological resources. Once a fire occurs, if it is not detected and controlled in time, it will cause immeasurable ecological damage, economic losses and even casualties. However, traditional forest firefighting is restricted by complex terrain and difficult to deploy manpower, making it difficult to achieve the goals of early detection, rapid response and effective containment of fires. In recent years, drone technology has emerged in the field of forest firefighting and has developed rapidly, showing many unique advantages in emergency warning and rapid fire extinguishing. Although some drone-based inventions have attempted to solve the problem of forest firefighting, there are still many limitations.

[0003] Chinese invention patent application CN118718292A discloses a forest fire extinguishing method and system based on drone clusters, which detects the initial fire through the monitoring front end, and then uses the reconnaissance drone to conduct secondary fire detection. After the fire is determined, the fire-fighting drone is remotely controlled to carry fire-fighting bombs to extinguish the fire. This application requires manual monitoring and fire-fighting drones are remotely controlled by humans, which is not convenient and intelligent enough.

[0004] Chinese invention patent application CN112132090A discloses a method for automatic detection and early warning of fireworks based on YOLOV3, which focuses on image processing and fire identification. It does not combine drone sensor data for comprehensive judgment, has a certain possibility of misjudgment, and cannot take fire-fighting measures in time.

[0005] Based on this, how to conduct dynamic inspections through drones, integrate sensor data to accurately identify forest fires and locate fire points, and how to use drones for rapid fire extinguishing response and re-ignition prevention, requires further research. Summary of the invention

[0006] The problem to be solved by the present invention is to provide a forest fire detection and extinguishing method and system, which can dynamically monitor the forest fire through unmanned aerial vehicles, satellite remote sensing and remote video, accurately identify the forest fire and locate the fire point, autonomously learn the optimal strategy through interaction with the environment, and perform resource allocation and task scheduling of unmanned aerial vehicles in complex environments.

[0007] The present invention adopts the following technical solution: a forest fire detection and extinguishing method, comprising the following steps: S1. Fire data collection and fire identification: Fire data is collected in real time through satellite remote sensing, remote video monitoring, and cruise drones, and transmitted to the command center to generate infrared thermal imaging images for fire judgment and fire grade classification; S2. Fire point location: The command center marks potential fire points based on infrared thermal imaging images, sorts potential fire points according to temperature values, obtains the geographic coordinates of the fire points from the GIS database, and updates in real time based on the movement and changes of the fire points in the continuous thermal imaging images, marking and locating new fire points. ; S3, calculation of fire spread trend: fire spread simulation is performed based on the Rothermel model, the fire spread trend is calculated and the fire spread boundary is updated; S4. Build a drone scheduling model and perform model training: Build a drone scheduling model based on the Markov decision process, and pre-train the drone scheduling model through the MAPPO multi-agent reinforcement learning algorithm; S5. Fire extinguishing by drone clusters: Fire-fighting drones are equipped with fire-fighting bombs. The command center uses the trained drone dispatching model to optimize the dispatching and decide the order of each drone’s throwing. With the shortest time as the optimization goal, the center collaboratively completes the investigation and extinguishing of all potential fire points. S6. Reignition prevention: After extinguishing the open fire, continue to patrol the fire scene with inspection drones to identify reignition points or smoke points.

[0008] Preferably, in step S1, satellite remote sensing is used to determine the boundaries of the fire and estimate the fire area; remote video monitoring is used to obtain forest scene images and videos in real time; there are multiple cruise drones to form a drone cluster, equipped with laser radar, infrared thermal imager and temperature, smoke and gas sensor devices, and the surface is detected by laser radar, the collected point cloud data is solved and orthographic photos are stitched, high-level precision DEM is extracted and topographic map is generated; the command center receives fire images and sensor data sent back by satellite remote sensing, remote video monitoring and cruise drones in real time, makes fire judgment and fire grade classification, displays them in the form of charts and heat maps, and issues review and fire extinguishing operation instructions after confirming the fire.

[0009] Preferably, in step S1, the command center performs fire situation judgment and fire grade classification in the following manner:

[0010] S1.1. Use the YOLOv5 method to detect fire on image data. Perform real-time detection through a single forward propagation network. Use data enhancement, pre-trained weights, and mixed precision training to strike a balance between speed and accuracy. Determine the confidence level of the image containing flames or smoke, with a value range of 0 to 1.

[0011] S1.2. Generate an infrared thermal image of the monitored area, divide the image into several temperature areas by setting a temperature threshold, and determine the high-temperature area; obtain the current fire situation by calculating the ratio of the high-temperature area to the entire image area; and preliminarily conclude that the fire is in a state of expansion or attenuation by comparing the changes in the area of ​​the high-temperature area in the same area over a period of time.

[0012] S1.3, process the field data collected by the drone's temperature, smoke, and gas sensors; when the ambient temperature exceeds the temperature alarm threshold, activate the fire alarm command, and continue to collect temperature data and transmit it to the command center. The command center pre-processes the temperature data over a period of time and takes the average as the regional temperature; The preprocessing adopts interpolation method to fill the missing data values, including linear interpolation and Lagrange interpolation, and estimates the value of the missing point based on the adjacent valid data. In principle, abnormal data is eliminated.

[0013] S1.4. For fire situation assessment, the command center uses a machine learning model to integrate data from different sensors with visual recognition results to comprehensively determine whether a fire has occurred. For fire level classification, the command center uses historical data to determine the fire risk of the area in the current season, and combines temperature, humidity, wind speed, and precipitation factors to set weights to obtain a fire score, and then classifies the current fire level based on the fire score.

[0014] Preferably, in step S3, the fire spread trend calculation includes the following sub-steps: S3.1. Collect geographic information and environmental data of the monitoring area as input of the Rothermel model: obtain topographic data corresponding to the geographic coordinates of the fire point from the GIS database, including slope and aspect information; obtain wind speed, wind direction data and fuel moisture information near the fire point from meteorological data; obtain fuel characteristic parameters by scanning the monitoring area through satellite remote sensing or drones, and establish an estimation model to estimate the fuel load by analyzing the relationship between the spectral characteristics of the image and the fuel load; The fuel characteristic parameters include: dried combustible content, dried particle density, surface area to volume ratio, combustibles, moisture content of combustibles, and extinguished moisture content of combustibles.

[0015] S3.2. Fire spread simulation based on the Rothermel model.

[0016] S3.3. Calculate the fire spread rate under initial windless and slopeless conditions: Based on the law of heat conduction and the input fuel characteristic parameters, under the ideal conditions of no wind and no slope, the fire spread formula is used to calculate the basic fire spread rate R1 when the fire is initially windless, no slope and the fuel humidity is a reference value (usually a standard dry state).

[0017] S3.4. Comprehensively consider the influence of wind speed, slope and fuel humidity, and calculate the actual spreading speed under the combined effect of wind speed and slope and the spreading speed against the wind: Based on the real-time collected wind speed and direction data, as well as the slope and aspect information corresponding to the fire point, the spread speed is adjusted to obtain the corrected fire spread speed, taking into account the wind's stretching and combustion-supporting effects on the flame, as well as the fuel accumulation and thermal radiation changes caused by the slope. The actual fire spread speed R2 under the combined effect of wind speed and slope, and the headwind spread speed R3 in the opposite direction of the fire head are calculated according to the correction formula.

[0018] S3.5. Update fire spread boundaries: Based on the fire spread speeds R1, R2, and R3, combined with the pre-set time step Based on the principle of kinematics, the fire spread position , the vector of velocity , predict the coordinates of the next fire feature point spread ; Connect the spread coordinate points with curves to obtain the new boundary of the fire spread, calculate the area of ​​the new area, update the fire spread trend in real time and visualize it to form a fire development situation map.

[0019] S3.6. Update fire spread speed and trend: Based on real-time updated infrared thermal imaging images, sensor data and meteorological data, the Rothermel model is adjusted to recalculate the fire spread speed and trend.

[0020] Preferably, in step S4, a drone scheduling model based on a Markov decision process is constructed as follows: First, let's describe the problem: The command center obtains the potential fire point sequence in real time through fire point positioning , assuming time Total M Fire Point , the position is recorded as ; Drone swarms have UAVs, each carrying The command center decides the order in which each drone will launch fire-extinguishing bombs, optimizing the shortest time possible to collaboratively complete the inspection and extinguishing of all potential fire points.

[0021] Construct a drone scheduling model based on Markov decision process, represented as a seven-tuple: ; in, is the state space; yes The joint action space of the agents; is the local observation space of the agent, It is an intelligent agent At a time step of Global state Local observation under is the state transition probability, indicating that given all agents in state Take joint action When, from the state Transfer to The probability of Take joint action The single-step reward after is the number of agents; is the discount factor, and its value range is [0,1].

[0022] Random Strategy It is a mapping from observation space to action space, expressed as: ; in, Representing an Agent A random strategy is used to give the probability distribution of an action under a specific observation.

[0023] Preferably, in the drone scheduling model, the state space The specific settings are as follows:

[0024] Each agent is assigned an independent state space. The state observation of an agent , defined as the position, linear velocity, rotation matrix and number of ammunition of the quadcopter, expressed as: ; in, , , , Respectively represent intelligent agents exist The position, linear velocity, rotation matrix and number of ammunition at the time.

[0025] No. Agent's environment observation , expressed as the relative position of other agents within the communication range: , Relative Position , the relative position of the lift and landing compartment , using the horizontal Euclidean distance approximation: ; Agent state environment observation , expressed as: ; in, For the fire scene Moment k indivual The flag indicating whether it is destroyed. for Moment k indivual priority.

[0026] t The state of the environment in the detection area at any moment is composed of all agent state environment observations, expressed as: ; Joint action space of agents , represented as a discrete action space , each agent action space is set the same , They respectively represent left, right, forward, backward, up, down, throw, and return.

[0027] The throwing action is allowed only when the distance to the target throwing point is less than the set threshold; the return action is only allowed when all tasks are completed and the potential fire points are gathered. Allows execution when it is empty and no longer updated for a period of time.

[0028] Preferably, a shared reward function is constructed in the drone scheduling model , maximize the next step while avoiding collision , shared reward function It is expressed as: ; in, For each simulation time step The reward function is: ; in, They represent obstacle avoidance reward, mission reward, all mission completion reward, distance reward, and return reward. is the weight coefficient, which is used to adjust the importance of each part of the reward, and its sum is 1.

[0029] Preferably, the obstacle avoidance reward , different rewards are given depending on whether the action taken causes the drone to collide with an obstacle: ; in, is the distance between the center of mass of the UAV and the obstacle, The maximum wheelbase of the drone.

[0030] Preferably, the task reward , associated with the task priority, and giving different rewards depending on whether the correct throw is completed: ; in, For the Step Completed gather, For the corresponding The priority of correspond The flag indicating whether it is destroyed.

[0031] Preferably, all task completion rewards : ; Only when all tasks are completed, the potential fire point set If it is empty and not updated for a period of time, it will be rewarded once.

[0032] Preferably, the distance reward , which is related to the shortest distance between each drone and the unfinished node. The coefficient adjusts the curvature of the reward function. As long as the target point is destroyed by any drone, all drones will receive a reward: ; For each unfinished ,and The distance between the nearest drones for: ; in, is the throwing height, ( )、( ) respectively represent the corresponding and the global coordinate position of the UAV.

[0033] Preferably, return flight rewards : ; When all mission points are completed, if the drone successfully flies to the nearest landing bay, a large positive reward will be given, otherwise a small negative reward will be given.

[0034] Preferably, in step S4, the MAPPO multi-agent reinforcement learning algorithm adopts a centralized training-decentralized execution framework to train two independent neural networks: an actor network and a central critic network; The actor network, represented as a Strategy Network , each agent has an independent actor network and makes independent decisions based on its own local information; The central critic network, represented as The state value function , all agents share a central critic network that evaluates the value of each agent by accessing the global state information of all agents.

[0035] The input feature of the actor network is normalized environmental information, and the output is the probability distribution of discrete actions; the central critic network inputs normalized centralized global state information and outputs state value estimation; The actor network and the central critic network both use multi-layer perceptrons to process observation information, with 256 neurons in the hidden layer, using ReLU activation function, and using Adam optimizer to maximize policy gradient and minimize value function loss for network update.

[0036] Preferably, in step S4, the drone scheduling model is pre-trained based on the MAPPO algorithm, including the following sub-steps: S4.1. The agent selects and executes actions in the environment according to the current strategy. The environment feeds back new states and rewards based on the agent's actions. The agent stores the information in the experience replay buffer and records the trajectory.

[0037] S4.2. Based on the collected trajectories, use the GAE method to estimate the advantage function , the estimation method is: S4.2.1. Calculate the TD error by Reflect the current state value function Estimated error: ; in, represents the discount factor; Indicates the termination sign; is the return reward after the agent takes the corresponding action, which is regularized by PopArt. .

[0038] S4.2.2. Calculate the GAE advantage estimate by introducing the attenuation factor TD error with different step numbers: ; in, Indicates the offset of the time step.

[0039] S4.3. For each actor training, calculate the following objective function: ; Among them, the parameters Represents actor network parameters, is the policy entropy hyperparameter, is the policy entropy, represents the clipped expected cumulative discounted reward.

[0040] S4.4. Maximization using gradient ascent : Calculate the policy loss gradient , and then use the stochastic gradient ascent method to update the actor network parameters , is the learning rate.

[0041] S4.5. For critic training, use gradient descent to minimize the following value function loss: ; in, represents a more accurate estimate of the expected future return given the current state, The function means limiting x to the range [a,b]. Represent the state value functions under the new and old parameters respectively, which are used to calculate the expected cumulative discounted reward; Indicated in strategy The expectation operator below.

[0042] S4.6. Repeat steps S4.1 to S4.5 until the strategy converges to the optimal solution or reaches the maximum number of training rounds.

[0043] The technical solution of the present invention also provides: a forest fire detection and extinguishing system for implementing any of the above methods, comprising: a fire identification module, a fire point location and fire spread trend calculation module, a fire extinguishing module, and a re-ignition prevention module; The fire identification module collects fire data in real time through satellite remote sensing, remote video monitoring, and cruise drones, transmits it to the command center, generates infrared thermal imaging images, and makes fire judgments and fire grade classifications; The fire point location and fire spread trend calculation module marks potential fire points based on infrared thermal imaging images, sorts potential fire points according to temperature values, accurately determines the geographical coordinates of fire points based on multi-source geographic information data, and updates in real time according to the movement and changes of fire points in continuous external thermal imaging images, marks and locates newly appeared fire points; and performs fire spread simulation based on the Rothermel model, calculates the fire spread trend and updates the fire spread boundary; The fire extinguishing module constructs a drone dispatching model based on the Markov decision process, and pre-trains the drone dispatching model through the MAPPO multi-agent reinforcement learning algorithm; the cruise drone is equipped with fire-extinguishing bombs, and the command center performs optimized dispatching based on the trained drone dispatching model, decides the throwing order of each drone, takes the shortest time as the optimization goal, and collaboratively completes the investigation and extinguishing of all potential fire points; After extinguishing the open fire, the re-ignition prevention module continues to patrol the fire scene through the inspection drone to identify the re-ignition point or smoke point.

[0044] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects: 1. The forest fire detection and extinguishing method of the present invention combines satellite remote sensing, remote video monitoring, and dynamic inspection by unmanned aerial vehicles to cover a larger detection range, reduce blind spots, and increase efficiency.

[0045] 2. The forest fire detection and extinguishing method of the present invention provides comprehensive early warning through fire identification and sensor data fusion, which is more timely, accurate and effective.

[0046] 3. In the UAV dispatching task of the method of the present invention, reinforcement learning can autonomously learn the optimal strategy through interaction with the environment, which is used to optimize the resource allocation and task dispatching of UAVs in dynamic and complex environments, making rescue faster, more optimized and safer. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flowchart of the forest fire detection and extinguishing method of the present invention; Figure 2 This is a schematic diagram of the application scenario of the MAPPO optimization strategy of the present invention; Figure 3 Schematic diagram of the MAPPO algorithm framework of the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the application is further elaborated in detail below in conjunction with the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments of other researchers in the field on this embodiment belong to the protection scope of the present invention. At the same time, for the step numbering in the embodiment of the present invention, it is only set for the convenience of explanation, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0049] In one embodiment of the present invention, a forest fire detection and extinguishing method, such as Figure 1 As shown, the following steps are included: S1. Fire data collection and fire identification: Fire data is collected in real time through satellite remote sensing, remote video monitoring, and cruise drones, and transmitted to the command center to generate infrared thermal imaging images for fire judgment and fire grade classification; S2. Fire point location: The command center marks potential fire points based on infrared thermal imaging images, sorts potential fire points according to temperature values, obtains the geographic coordinates of the fire points from the GIS database, and updates in real time based on the movement and changes of the fire points in the continuous thermal imaging images, marking and locating new fire points; S3, calculation of fire spread trend: fire spread simulation is performed based on the Rothermel model, the fire spread trend is calculated and the fire spread boundary is updated; S4. Build a drone scheduling model and perform model training: Build a drone scheduling model based on the Markov decision process, and pre-train the drone scheduling model through the MAPPO multi-agent reinforcement learning algorithm; S5. Fire extinguishing by drone clusters: Fire-fighting drones are equipped with fire-fighting bombs. The command center uses the trained drone dispatching model to optimize the dispatching and decide the order of each drone’s throwing. With the shortest time as the optimization goal, the center collaboratively completes the investigation and extinguishing of all potential fire points. S6. Reignition prevention: After extinguishing the open fire, continue to patrol the fire scene with inspection drones to identify reignition points or smoke points.

[0050] Specifically, in step S1, fire data collection and fire identification include the following sub-steps: 1. Combine satellite remote sensing, remote video monitoring and drone patrol to collect fire data.

[0051] Satellite remote sensing technology can cover a wide geographical area and has significant advantages in large-scale, cross-regional forest fire monitoring. It can more accurately determine the boundaries of the fire and estimate the fire area, providing important reference for subsequent fire fighting and rescue work.

[0052] For some small-scale fires or early-stage fires, satellite remote sensing technology may not be able to accurately identify and locate them, and may easily result in missed reports or false reports. Therefore, this embodiment sets up monitoring points in key forest areas or areas with high fire incidence, and combines the remote video monitoring system of forest control to obtain real-time images and videos of the forest scene and transmit them back to the command center.

[0053] Due to factors such as terrain and tree obstruction, some areas may not be monitored, thus forming blind spots. Therefore, this embodiment dispatches multiple patrol drone clusters to conduct fire inspections in the planned forest area. The drones are equipped with temperature sensors, smoke alarm sensors, infrared thermal imaging sensors, gas sensors and other sensor devices, which transmit relevant images and sensor collection data to the command center in real time for fire judgment and fire level classification.

[0054] The drone cluster is also equipped with a laser radar to detect the ground surface. After solving the point cloud data and stitching the orthographic photos, it extracts high-level precision DEM and generates a topographic map.

[0055] 2. The command center combines data from all parties to make a fire situation assessment and classify the fire level.

[0056] The command center uses data from all parties as a criterion for fire situation judgment, sets their respective weights, and obtains the weighted fire score. It then divides the fire level according to the score. The data used include: 1) Image data: Use YOLOv5 to detect fire on image data and determine the confidence level of flames or smoke in the image, with a value range of 0 to 1.

[0057] YOLOv5 is an efficient and flexible target detection model. In this embodiment, based on the YOLOv5 model, real-time detection is achieved through a single forward propagation network, and a good balance is achieved between speed and accuracy through technologies such as data enhancement, pre-trained weights, and mixed precision training.

[0058] 2) Infrared thermal imaging: By setting the temperature threshold, the image is divided into several temperature areas, and the ratio of the high-temperature area to the entire image area is calculated to obtain the current fire situation; further, by comparing the changes in the high-temperature area in the same area over a period of time, it is preliminarily concluded that the fire is in a state of expansion or attenuation.

[0059] 3) Sensor data: Data collected by temperature, smoke, gas and other sensor devices mainly comes from on-site collection by drones.

[0060] Taking the temperature sensor as an example, when the ambient temperature exceeds the temperature alarm threshold, the temperature sensor activates the fire alarm and continuously collects the temperature around the area and transmits it to the command center; the command center pre-processes the collected temperature during this period and takes the average as the temperature of the area.

[0061] The specific preprocessing method is: if data is missing, use interpolation methods to fill in the missing values, such as linear interpolation, Lagrange interpolation, etc., and estimate the value of the missing point based on the valid data before and after; if data is abnormal, use 3 Principle elimination.

[0062] 4) Forest fire weather level on the day (referred to as fire risk level): released by the China Meteorological Administration and the Forestry Bureau, mainly calculated based on factors such as temperature, humidity, wind speed, and precipitation.

[0063] In this embodiment, the fire danger levels 1 to 5 are quantified as 0, 20, 60, 80, and 100. The command center uses machine learning models (such as SVM, random forest, and neural network) to fuse the data from different sensors with the visual recognition results to comprehensively determine whether a fire has occurred.

[0064] 5) Historical data: The fire risk level of the area in the current season obtained through historical data.

[0065] 3. After confirming the fire, inform the inspection personnel to review and start fire-fighting operations. The fire data and analysis results collected by the drone are displayed in the form of charts, heat maps, etc., which is convenient for relevant departments to monitor and make decisions.

[0066] Specifically, in step S2, the fire point positioning includes the following sub-steps: 1. Thermal imaging data collection and transmission Based on the real-time thermal imaging data collected by infrared thermal imaging equipment, including remote video monitoring and inspection drones, a stable and high-speed data transmission link is established by combining multiple technologies such as 5G network and satellite communication to ensure that the image data collected by the thermal imaging equipment is transmitted to the command center in real time, avoiding the impact of transmission delays or interruptions on fire point positioning and trend judgment.

[0067] 2. Fire point screening and positioning in the command center After receiving the infrared thermal imaging images, the command center uses professional image processing software and algorithms to quickly analyze the images. By setting the temperature threshold, areas with a certain range of temperatures higher than the normal ambient temperature are marked as potential fire points, and the fire points are prioritized according to the temperature value. Further combined with multi-source geographic information data, the geographic coordinates of the fire point are accurately determined, and the error is controlled within a very small range for the precise implementation of subsequent firefighting operations.

[0068] 3. Real-time fire location update The infrared thermal imaging monitoring equipment continuously scans the monitoring area, recollects image data at fixed time intervals (such as every 30 seconds) and transmits it to the command center. The command center automatically compares the thermal images of the previous and subsequent periods, and updates the geographical coordinates of the fire points in real time based on the movement and changes of the fire points in the images, combined with the geographic information system (GIS) technology. At the same time, the newly appeared fire points are marked and located in a timely manner to ensure comprehensive control of the dynamics of the fire scene.

[0069] Specifically, in step S3, the fire spread trend calculation includes the following sub-steps: 1. Geographic information and environmental data collection Utilize high-precision geographic surveying and mapping technology to obtain detailed topographic data in the monitoring area in advance, including slope and aspect information, and integrate it into the GIS database.

[0070] For the location of the fire point, after precise positioning through infrared thermal imaging combined with GIS technology, the corresponding slope and slope direction information is retrieved from the GIS database in real time and fed back to the input data interface of the subsequent fire spread model, providing basic terrain data for accurate simulation of fire spread.

[0071] The command center has established a close data sharing cooperation mechanism with the local meteorological bureau to obtain real-time wind speed, wind direction data and fuel humidity-related information near the fire point. These data will be used as key input parameters for subsequent fire spread calculations.

[0072] In addition, satellite remote sensing is used to scan the monitoring area regularly, and professional data processing algorithms are used to estimate the fuel load, including parameters such as dried combustible content, dried particle density, surface area to volume ratio, combustibles and moisture content of combustibles, and moisture content of combustibles before extinguishing, to provide comprehensive and accurate fuel property data for the fire spread model.

[0073] 2. Fire spread simulation based on Rothermel model In this embodiment, the Rothermel model takes into account factors such as slope and aspect, fuel bed and forest density, obeys various heat conduction laws, and has universal application. The expression is as follows: ; Where, R is the forest fire spread speed, unit (m / min); is the flame zone reaction intensity, in units of ; is the forest fire spread rate, is the wind speed correction factor, is the slope correction factor, is the effective heat coefficient, both are dimensionless coefficients; is the density of the combustible material, in units ; The heat required to ignite unit mass of combustible material, unit (kJ / kg).

[0074] The specific calculation steps of the Rothermel model are as follows: 1) Input data The acquired fire point location information, wind speed, wind direction, dried combustible content, dried particle density, surface area to volume ratio, combustibles, moisture content of combustibles, moisture content of combustibles before extinguishing, and other parameters, as well as parameters ( , ) is obtained by fitting the empirical formula from historical data and accurately input into the model calculation interface in sequence according to the requirements of the Rothermel model.

[0075] 2) Calculate the fire spread speed under initial windless and slopeless conditions Based on the law of heat conduction and the input fuel characteristic parameters, under the ideal conditions of no wind and no slope, the built-in fire spread formula is used to perform complex operations on various parameters to calculate the fire spread speed under the initial conditions of no wind and no slope. This speed serves as the basic data for subsequent consideration of the influence of environmental factors.

[0076] 3) Calculate the actual spreading speed under the combined effect of wind speed and slope and the speed against the wind We further introduced real-time collected wind speed and direction data as well as the slope and aspect information corresponding to the fire point location, and comprehensively considered factors such as the wind's stretching and combustion-supporting effects on the flame, as well as the fuel accumulation and thermal radiation changes caused by the slope. We then used a specific correction formula to accurately adjust the spread speed, and calculated the actual spread speed of the fire under the combined effect of wind speed and slope, as well as the upwind spread speed in the opposite direction of the fire head.

[0077] A specific correction formula accurately adjusts the spread rate, taking into account the fire spread rate after the wind speed, slope, and fuel moisture are affected: ; in, U Indicates wind speed, usually the average wind speed at a certain height (such as 10 meters) above the ground; S Slope, usually expressed as the tangent of the slope angle, which is the ratio of vertical height to horizontal distance; M Indicates fuel moisture, expressed as a percentage; It is an empirical coefficient, calibrated according to experimental data or historical fire data.

[0078] According to the above correction formula, the actual spreading speed R2 of the fire under the combined effect of wind speed and slope, and the headwind spreading speed R3 in the opposite direction of the fire head are calculated respectively.

[0079] 4) Update the fire spread boundary Based on the calculated fire spread speed under different conditions, combined with the pre-set time step (such as one calculation step every 15 minutes), the kinematic principle is used to calculate the fire spread position. , the vector of velocity , predict the coordinates of the next fire feature point (flame front, high temperature core area boundary and other key locations) : ; ; The coordinates of these spreads The points are connected with curves to obtain the new boundary of the fire spread during this period and the area of ​​the new region; at the same time, the fire spread trend information is updated and visualized in real time to provide an intuitive and dynamic fire development situation map.

[0080] As the fire develops and evolves, thermal images, sensor data, and meteorological data are continuously updated. The system continuously feeds real-time new data into the Rothermel model. The model automatically adjusts parameters based on the new information and recalculates the fire spread rate and trend to ensure that the prediction results are always close to the actual situation at the fire scene.

[0081] Through the above-mentioned fire location and fire spread trend calculation, accurate and timely decision-making basis is provided for fire-fighting operations, such as providing reliable action guidance for the dispatching task of fire-fighting drones to drop fire-fighting bombs, ensuring the efficient allocation and use of fire-fighting resources.

[0082] Specifically, in step S4, a drone scheduling model is constructed and model training is performed, including the following sub-steps: Problem description: The command center obtains a dynamic and updateable potential fire point sequence in real time through fire point positioning , assuming time Total M Fire Point , whose position is recorded as .

[0083] In this embodiment, the existing problem is: The number of drones carried is To deploy fire-extinguishing bombs, the system needs to independently decide the order in which each drone will throw the bombs, with the shortest time as the optimization goal, and collaboratively complete the inspection and extinguishing of all potential fire points.

[0084] The multi-UAV scheduling problem involves multiple complex task point allocations. The dynamic update of task points further increases the difficulty of the problem. Traditional methods are difficult to deal with in real time and effectively. This embodiment adopts the reinforcement learning method to learn the optimal strategy through the interaction between the intelligent agent and the environment, providing a good solution to this problem.

[0085] Specifically, this embodiment uses the Markov decision process (DEC-POMDP) ​​to describe the decision model of dynamic scheduling and control of the drone cluster, and defines the model as a seven-tuple: ; in, is the state space; yes The joint action space of the agents; is the agent i at time step Global state Local observation under is the state transition probability, indicating that given all agents in state Take joint action When, from the state Transfer to The probability of Take joint action The reward for the next step is used by the reward function to calculate the reward in state Take joint action Rewards received; is the number of agents; is a discount factor, ranging from [0,1], which decays with the number of training steps.

[0086] Random Strategy It is a mapping from observation space to action space, giving the probability distribution of an action under a specific observation, denoted as ;in, Representing an Agent A random strategy is used to give the probability distribution of an action under a specific observation.

[0087] 1. State Observation Space

[0088] Each agent is assigned an independent state space. The state observation of each agent is defined as the position, linear velocity, rotation matrix and number of ammunition of the quadcopter, expressed as: ; in, , , , Respectively represent intelligent agents exist The position, linear velocity, rotation matrix and number of ammunition at the time.

[0089] No. The environment observation of an agent is the communication with other agents within the communication range. and the relative position of the landing chamber (fixed position), where the agent is The relative distances to all landing compartments are expressed using horizontal Euclidean distance: ; Agent state environment observation , expressed as: ; in, For the fire scene Moment k indivual The flag indicating whether it is destroyed. for Moment k indivual priority.

[0090] The state of the environment in the detection area at time t is composed of all agent state environment observations, expressed as: .

[0091] 2. Joint Action Space

[0092] Joint Action Space , represented as a discrete action space , each agent action space is set the same , They respectively represent left, right, forward, backward, up, down, throw, and return.

[0093] The throwing action is allowed only when the distance to the target is less than a certain value, otherwise the action is prohibited; similarly, the return action is only allowed when all tasks are completed and the potential fire points are gathered. When it is empty and no longer updated for a period of time, execution is allowed; otherwise, the action is prohibited.

[0094] 3. Shared Reward Function

[0095] The goal is to maximize the propulsion of the next ; For a fully cooperative problem, the reward function for all agents is the same, and the reward function at each simulation time step t is defined as: ; in, is the weight coefficient, which is used to adjust the importance of each part of the reward, and its sum is 1; They represent obstacle avoidance reward, mission reward, reward for completing all missions, distance reward, and return reward respectively.

[0096] Therefore, the shared reward function is: ; (1) Obstacle avoidance reward: ; in, is the distance between the center of mass of the UAV and the obstacle, The maximum wheelbase of the drone.

[0097] If the action taken causes the drone to collide with an obstacle, a large negative reward is given; if it is relatively close to the obstacle, a certain negative reward is given; otherwise, a small positive reward is given.

[0098] (2) Mission rewards: ; in, For the Step Completed gather, For the corresponding priority.

[0099] Once in the unfinished If you complete the throw at the right time, you will get a positive reward, and the reward value is linked to the task priority; otherwise, an incorrect throw will get a huge negative reward.

[0100] In the actual process, the drone checks the current position and Check whether the horizontal Euclidean distance is within an acceptable range, and review the potential fire point. After confirming it is a fire point, throw fire extinguishing bombs.

[0101] (3) One-time reward for completing the task: ; Only when the potential fire point set If it is empty and there is no update for a period of time, it will be rewarded once.

[0102] (4) Distance Reward: For each unfinished , and the distance between it and the nearest drone is: ; in, is the throwing height; )、( ) respectively represent the corresponding and the global coordinate position of the UAV.

[0103] Construct the following distance function, which is related to the shortest distance of each drone to the unfinished node. As long as the target point is destroyed by any drone, all drones will receive a positive reward to ensure that there is no competition between multiple drones: ; In addition, this item serves as a guide item to make the reward more intensive, which can be adjusted The coefficient adjusts the curvature of the reward function.

[0104] (5) Return flight rewards: ; When all mission points are completed, if the drone successfully flies to the nearest landing bay, a large positive reward will be given, otherwise a small negative reward will be given.

[0105] Furthermore, the drone scheduling model is pre-trained through the MAPPO multi-agent reinforcement learning algorithm. Different batches of drones are pre-trained to cope with fire extinguishing tasks under various fire conditions. The command center dispatches a total of N fire-fighting drones from the landing bay closest to the fire scene according to the fire level, and calls the pre-trained model.

[0106] In this embodiment, Figure 2 As shown, each drone is set to carry a maximum of four fire-extinguishing bombs. Each drone will achieve autonomous obstacle avoidance, assist in extinguishing dynamically updated fire points, and carry out anti-cabin operations nearby.

[0107] In particular, PPO (Proximal Policy Optimization) is an on-policy reinforcement learning algorithm that is improved on the basis of the policy gradient algorithm. It has achieved strong final returns and sample efficiency in various cooperative multi-agent challenges and has been widely used.

[0108] In this embodiment, the MAPPO (PPO in Multi-Agent Settings) multi-agent reinforcement learning algorithm is an extension of the PPO algorithm in the multi-agent field. Figure 3 As shown in the figure, the centralized learning and decentralized execution (CLDE) framework is used to train two independent neural networks: one with parameters Strategy Network (also called an actor), and a The state value function (also called critic).

[0109] All agents share a central critic network, which evaluates the value of each agent more accurately by accessing the global state information of all agents. At the same time, each agent has an independent actor network and can make independent decisions based on its own local information, improving the flexibility and robustness of the system.

[0110] In this embodiment, the specific training process of the drone scheduling model is as follows: The agent selects actions and executes them in the environment according to the current strategy. The environment feeds back new states and rewards based on the agent's actions. The agent stores this information in the experience replay buffer and records the trajectory. MAPPO maximizes the cumulative reward based on the policy gradient and uses the stored experience data to update the value network of the agent.

[0111] For each actor training, maximize the objective function: ; Among them, the parameters Represents actor network parameters, is the policy entropy hyperparameter.

[0112] Represents policy entropy: ; in, Indicates the current strategy The probability of taking action a in state s;

[0113] represents the clipped expected cumulative discounted reward:

[0114] ; in, Indicated in strategy The expectation operator under is the crop factor, the ratio function Represents the action probability ratio of the new and old strategies at time t.

[0115] Maximize using gradient ascent , calculate the policy loss gradient , and then use the stochastic gradient ascent method to update the actor network parameters : ; in, is the learning rate; These are the actor network parameters before updating.

[0116] For advantage function estimation, the GAE (Generalized Advantage Estimation) method is used for estimation and the PopArt (Preserving Outputs Precisely, while Adaptively Rescaling Targets) method is used for regularization. The calculation steps are as follows: 1) Calculate the TD (Temporal Difference) error: Reflects the current state value function Estimated error: ; in, represents the discount factor; Indicates the termination sign; is the return reward after the agent takes the corresponding action, which is regularized by PopArt. ;

[0117] 2) Calculate the GAE advantage estimate: by introducing a decay factor To combine the TD errors of different steps: ; when When , it degenerates into a single-step TD estimation; when , which is equivalent to Monte Carlo estimation.

[0118] In actual calculation, since a trajectory is a finite time step [0,…,T], the above formula can be approximated as: ; For critic training, minimize the value function loss: ; in, represents a more accurate estimate of the expected future return given the current state, Function means x Restricted to [ a , b ] range, Represent the state value function under the new and old parameters, respectively, which is used to calculate the expected cumulative discounted reward, Represents the crop factor.

[0119] In the actual execution process, the agents call their own optimal strategies to interact with the environment, repeating the above environmental interaction and learning steps until they learn the optimal strategy that can maximize the long-term cumulative reward. The specific algorithm steps are as follows:

[0120] 1) Initialize weights and And orthogonal initialization, initialize the experience replay pool ;

[0121] 2) Initialize each agent state Start the iteration of training rounds;

[0122] 3) Each agent has its own strategy function Interact with the environment, collect trajectories, and record experience data at each time step t And store it in the experience replay pool;

[0123] 4) Based on the collected trajectories, use the GAE method to estimate the advantage function ;

[0124] 5) Randomly sample a batch of experience data from the experience replay pool for calculation in steps 6 and 7 below;

[0125] 6) Using advantage estimation, calculate the policy function loss , update the actor network parameters using the gradient ascent method ;

[0126] 7) Calculate the value function loss , use the gradient descent method to update the critic network parameters ;

[0127] 8) Repeat this process for multiple iterations until the strategy converges to the optimal solution or reaches the maximum number of training rounds.

[0128] In the specific training process of this embodiment, the policy network inputs the environment information after feature normalization and outputs the probability distribution of discrete actions. The value network inputs the centralized global state information after normalization and outputs the state value estimate. Multilayer perceptron (MLP) is used to process the observation information, the number of hidden layer neurons is 256, and the ReLU activation function is used. The Adam optimizer is used to maximize the policy gradient and minimize the value function loss to achieve network update.

[0129] Furthermore, model evaluation: the model is evaluated using a test environment, which differs from the training environment in some parameters or scenario settings, but maintains the same basic structure and rules. Model evaluation includes the cumulative reward in one evaluation round, the calculation of the average reward and its standard deviation over multiple evaluation rounds, and the success rate of the agent in completing the task.

[0130] Furthermore, model parameter adjustment: If the model performance is not ideal, it is necessary to further adjust hyperparameters such as learning rate, discount factor, discount factor of advantage estimate, entropy coefficient, clipping range, training batch size, exploration rate, or adjust network structure parameters such as the number of neurons and layers, and activation function type.

[0131] Specifically, in step S5, the fire of the drone cluster is extinguished as follows:

[0132] The cruise drones are equipped with fire-extinguishing bombs. The command center optimizes the scheduling through the trained drone scheduling model, independently decides the throwing order of each drone, takes the shortest time as the optimization goal, and collaboratively completes the inspection and destruction of all potential fire points.

[0133] Specifically, in step S6, the method for preventing re-ignition is as follows:

[0134] After extinguishing the open fire, the inspection drone patrols the fire scene to identify the re-ignition point or smoke point to prevent a second large-scale fire. After receiving the re-ignition point information from the inspection drone, the command center immediately activates the emergency response plan. On the one hand, according to the location of the re-ignition point, the nearest firefighting force is quickly dispatched, such as ground personnel carrying fire extinguishers, drones equipped with fire extinguishing bombs, etc., to go to the rescue to ensure that the fire is controlled in the early stage of the re-ignition; on the other hand, a secondary alert is carried out around the re-ignition area to evacuate people who may be affected to prevent a second large-scale fire from causing greater losses.

[0135] This embodiment also includes: a forest fire detection and extinguishing system for implementing the above method, including: a fire identification module, a fire point location and fire spread trend calculation module, a fire extinguishing module, and a re-ignition prevention module;

[0136] The fire identification module collects fire data in real time through satellite remote sensing, remote video monitoring, and cruise drones, transmits it to the command center, generates infrared thermal images, and makes fire judgments and fire grade classifications;

[0137] The fire point location and fire spread trend calculation module marks potential fire points based on infrared thermal imaging images, sorts potential fire points according to temperature values, and accurately determines the geographical coordinates of fire points based on multi-source geographic information data. It updates in real time based on the movement and changes of fire points in continuous external thermal imaging images, marks and locates new fire points; and simulates fire spread based on the Rothermel model, calculates fire spread trends and updates fire spread boundaries;

[0138] The fire extinguishing module builds a drone dispatching model based on the Markov decision process and pre-trains the drone dispatching model through the MAPPO algorithm. The command center optimizes the dispatching of fire-extinguishing bombs carried out by the cruise drone based on the trained drone dispatching model, decides the order of each drone’s throwing, takes the shortest time as the optimization goal, and collaboratively completes the investigation and extinguishing of all potential fire points.

[0139] The re-ignition prevention module, after extinguishing the open fire, continues to patrol the fire scene through inspection drones to identify the re-ignition point or smoke point.

[0140] In summary, the forest fire detection and extinguishing method of the present invention uses multiple patrol drones to conduct fire inspections on the planned forest area, transmit images and airborne sensor data to the command center in real time, use artificial intelligence methods to identify fire conditions and combine data from all parties to make fire judgments and fire level classifications, and inform relevant personnel to review and start operations after confirming the fire conditions. The command center screens fire points based on thermal imaging images, updates the location of fire points in real time, and judges the spread trend of the fire scene based on the returned data, providing a basis for the dispatching task of the fire-fighting drone to throw fire-fighting bombs. The MAPPO multi-agent reinforcement learning algorithm is used to train the optimal dynamic fire point collaborative extinguishing control strategy to achieve efficient collaborative fire extinguishing, dynamic environmental adaptation and resource optimization allocation, and maintain robustness when some agents fail. At the same time, after extinguishing the open fire, the patrol drone patrols the fire scene to identify the re-ignition point or smoke point to prevent a second large-scale fire.

[0141] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A forest fire detection and extinguishing method, characterized in that: The steps include: S1. Fire data collection and fire identification: Fire data is collected in real time through satellite remote sensing, remote video monitoring, and cruise drones, and transmitted to the command center to generate infrared thermal imaging images for fire judgment and fire grade classification; S2. Fire point location: The command center marks potential fire points based on infrared thermal imaging images, sorts potential fire points according to temperature values, obtains the geographic coordinates of the fire points from the GIS database, and updates in real time based on the movement and changes of the fire points in the continuous thermal imaging images, marking and locating new fire points; S3, fire spread trend calculation: fire spread simulation is performed based on the Rothermel model, the fire spread trend is calculated and the fire spread boundary is updated; S4. Build a drone scheduling model and perform model training: Build a drone scheduling model based on the Markov decision process and pre-train the drone scheduling model through the MAPPO multi-agent reinforcement learning algorithm; S5. Fire extinguishing by drone clusters: Fire-fighting drones are equipped with fire-fighting bombs. The command center uses the trained drone dispatching model to optimize the dispatching and decide the order of each drone’s throwing. With the shortest time as the optimization goal, the center collaboratively completes the investigation and extinguishing of all potential fire points. S6. Reignition prevention: After extinguishing the open fire, continue to patrol the fire scene with inspection drones to identify reignition points or smoke points.

2. The forest fire detection and extinguishing method according to claim 1, characterized in that: In step S1, the satellite remote sensing is used to determine the boundaries of the fire and estimate the fire area; the remote video monitoring is used to obtain real-time forest scene images and videos; there are multiple cruise drones to form a drone cluster, equipped with laser radar, infrared thermal imager and temperature, smoke and gas sensors, and the surface is detected by laser radar, the collected point cloud data is solved and orthographically stitched, high-level precision DEM is extracted and a topographic map is generated.

3. The forest fire detection and extinguishing method according to claim 2, characterized in that: In step S1, the command center receives fire images and sensor data sent back by satellite remote sensing, remote video monitoring, and cruise drones in real time to make fire judgments and classify fire levels in the following manner: S1.

1. Use the YOLOv5 method to perform real-time fire detection on image data through a single forward propagation network, balance speed and accuracy through data enhancement, pre-trained weights, and mixed precision training to determine the confidence that the image contains flames or smoke; S1.

2. Generate infrared thermal imaging images of the monitoring area, divide the image into several temperature areas by setting temperature thresholds, and determine the high-temperature area; calculate the ratio of the high-temperature area to the entire image area to obtain the current fire situation; and preliminarily judge whether the fire is expanding or attenuating by comparing the changes in the area of ​​the high-temperature area within a certain period of time in the monitoring area; S1.3, continuously collect the temperature, smoke, and gas sensor data of the drone and transmit it to the command center. The command center pre-processes the temperature data over a period of time and takes the average as the regional temperature. When the regional temperature exceeds the temperature alarm threshold, the fire alarm command is activated; S1.

4. The command center uses a machine learning model to integrate different sensor data with visual recognition results to comprehensively determine whether a fire has occurred. The command center also uses historical data to determine the fire risk in the monitoring area in the current season, combined with temperature, humidity, wind speed, and precipitation factors, to set weights to obtain a fire score, and then divides the current fire level according to the fire score.

4. The forest fire detection and extinguishing method according to claim 2, characterized in that: In step S3, the fire spread trend calculation includes the following sub-steps: S3.

1. Collect geographic information and environmental data of the monitoring area as input to the Rothermel model: Obtain topographic data corresponding to the geographic coordinates of the fire point from the GIS database, including slope and aspect information; obtain wind speed and direction data near the fire point and fuel moisture information from meteorological data; obtain fuel characteristic parameters through satellite remote sensing or drone scanning of the monitoring area, and estimate fuel load by analyzing the relationship between the spectral characteristics of the image and the fuel load; The fuel characteristic parameters include: dried combustible content, dried particle density, surface area to volume ratio, combustibles, moisture content of combustibles, and extinguished moisture content of combustibles; S3.

2. Fire spread simulation is performed based on the Rothermel model. The fire spread formula is: ； In the formula, R is the speed at which forest fire spreads; is the reaction intensity of the flame zone; is the forest fire spread rate; is the wind speed correction factor; is the slope correction factor; is the effective thermal coefficient; is the density of the combustible; The amount of heat required to ignite a unit mass of combustible material; S3.

3. Calculate the fire spread speed under the initial conditions of no wind and no slope: Based on the law of heat conduction and the input fuel characteristic parameters, the fire spread formula is used to calculate the basic fire spread speed R1 when there is no wind, no slope and the fuel humidity is the reference value under the assumption of no wind and no slope. S3.

4. Calculate the actual spreading speed under the combined effect of wind speed and slope and the upwind spreading speed: Based on the real-time wind speed and direction data and the slope and slope direction information corresponding to the fire point, the spread speed is corrected by considering the wind's stretching and combustion-supporting effects on the flame, as well as the fuel accumulation and thermal radiation changes caused by the slope, and the corrected fire spread speed is obtained. The correction formula is: ； in, U Indicates wind speed, using the average wind speed at a preset height above the ground; S Indicates the slope, expressed as the tangent value of the slope angle; M Indicates fuel moisture, expressed as a percentage; It is an empirical coefficient, which is calibrated according to experimental data or historical fire data; According to the correction formula, the actual fire spreading speed R2 under the combined effect of wind speed and slope, as well as the headwind spreading speed R3 in the opposite direction of the fire head are calculated respectively; S3.

5. Update fire spread boundaries: Based on the fire spread speed R1, R2, R3, combined with the time step , based on the principle of kinematics, according to the fire spread position , velocity vector , predict the coordinates of the next fire feature point spread point : ； ； Connect the spread coordinate points with curves to obtain the new boundary of fire spread, calculate the area of ​​the new area, update the fire spread trend in real time and visualize it to form a fire development situation map; S3.

6. Update fire spread speed and trend: Based on the real-time updated infrared thermal imaging images, sensor data and meteorological data, repeat steps S3.1 to S3.5, adjust through the Rothermel model, and update the calculation of fire spread speed and trend.

5. The forest fire detection and extinguishing method according to claim 4, characterized in that: In step S4, a UAV scheduling model based on Markov decision process is constructed as follows: Problem description: The command center obtains the potential fire point sequence in real time through fire point positioning , assuming time Collection M Fire Point , the position is recorded as ; Drone swarms have UAVs, each carrying The command center decides the order of each drone's throwing of fire extinguishing bombs, with the shortest time as the optimization goal, to collaboratively complete the investigation and destruction of all potential fire points; Construct a drone scheduling model based on Markov decision process, represented as a seven-tuple: ； in, is the state space; is the number of agents; yes The joint action space of the agents; is the local observation space of the agent, It is an intelligent agent At a time step of Global state Local observation under is the state transition probability, indicating that all given agents are in state Take joint action When, from the state Transfer to The probability of Take joint action The single-step reward after is the discount factor; Random Strategy It is a mapping from observation space to action space, expressed as: ； in, Representing an Agent The random strategy of .

6. The forest fire detection and extinguishing method according to claim 5, characterized in that: In the drone scheduling model, each agent is assigned an independent state space. The state observation of an agent Defined as the position, linear velocity, rotation matrix and number of ammunition of the quadcopter drone, expressed as: ； in, , , , Respectively represent intelligent agents exist t The position, linear velocity, rotation matrix and number of ammunition at the time; No. Agent's environment observation Use horizontal Euclidean distance approximation: ； in, Indicates The relative positions of the agents; Indicates the fire point The relative position of Indicates the relative position of the landing compartment; Agent state environment observation It is expressed as: ； in, For the fire scene Moment k indivual The flag indicating whether it is destroyed. for Moment k indivual Priority; t The state of the environment in the detection area at any moment is composed of all agent state environment observations, expressed as: ； Joint action space of agents , represented as a discrete action space , the action space of each agent is set the same, expressed as , They respectively represent left, right, forward, backward, up, down, throw, and return.

7. The forest fire detection and extinguishing method according to claim 6, characterized in that: In the drone scheduling model, a shared reward function is constructed , maximize the next step while avoiding collision : ； in, For each simulation time step t, the reward function is: ； in, They represent obstacle avoidance reward, mission reward, all mission completion reward, distance reward, and return reward. is the weight coefficient, which is used to adjust the importance of each part of the reward; Obstacle Avoidance Reward , different rewards are given depending on whether the action taken causes the drone to collide with an obstacle: ； in, is the distance between the center of mass of the UAV and the obstacle, The maximum wheelbase of the drone; Mission Rewards , associated with the task priority, and giving different rewards depending on whether the correct throw is completed: ； in, For the Step Completed gather, For the corresponding The priority of correspond The flag indicating whether it is destroyed; All mission completion rewards , only when all tasks are completed, the set of potential fire points If it is empty and no longer updated for a period of time, a reward will be given once: ； Distance Rewards , which is related to the shortest distance between each drone and the unfinished node. The coefficient adjusts the curvature of the reward function. As long as the target point is destroyed by any drone, all drones will receive a reward: ； For each unfinished ,and The distance between the nearest drones for: ； in, is the throwing height, ( )、( ) respectively represent the corresponding and the global coordinate position of the UAV; Return flight rewards : ； When all mission points are completed, the drone will be given a positive reward if it successfully flies to the nearest landing bay, otherwise it will be given a negative reward.

8. The forest fire detection and extinguishing method according to claim 5, characterized in that: In step S4, the MAPPO multi-agent reinforcement learning algorithm adopts a centralized training-distributed execution framework to train two independent neural networks: an actor network and a central critic network; The actor network, represented as a Strategy Network , each agent has an independent actor network and makes independent decisions based on its own local information; the input features of the actor network are normalized environmental information, and the output is the probability distribution of discrete actions; The central critic network, represented as The state value function , all agents share a central critic network, which evaluates the value of each agent by accessing the global state information of all agents; the input of the central critic network is the normalized centralized global state information, and the output is the state value estimate; The actor network and the central critic network process observation information through multi-layer perceptrons respectively, the number of hidden layer neurons is 256, the ReLU activation function is used, and the Adam optimizer is used to maximize the policy gradient and minimize the value function loss for network update.

9. The forest fire detection and extinguishing method according to claim 8, characterized in that: In step S4, the drone scheduling model is pre-trained based on the MAPPO algorithm, including the following sub-steps: S4.

1. The agent selects and executes actions in the environment according to the current strategy. The environment feeds back new states and rewards based on the agent's actions. The agent stores the information in the experience replay buffer and records the trajectory. S4.

2. Based on the collected trajectories, use the GAE method to estimate the advantage function , the estimation method is: Calculate the TD error by Reflect the current state value function Estimated error: ； in, represents the discount factor; Indicates the termination sign; is the return reward after the agent takes the corresponding action, which is regularized by PopArt. ; Calculate the GAE advantage estimate by introducing a decay factor TD error with different step numbers: ； in, Indicates the offset of the time step; S4.

3. Train each actor network and calculate the objective function: ； Among them, the parameters Represents actor network parameters, is the policy entropy hyperparameter, is the policy entropy, represents the clipped expected cumulative discounted reward; S4.

4. Maximization using gradient ascent , calculate the policy loss gradient , and then use the stochastic gradient ascent method to update the actor network parameters : ； in, is the learning rate; is the actor network parameter before updating; S4.

5. Train the central critic network and use gradient descent to minimize the value function loss: ； in, represents the estimate of the expected future return given the current state, Function means x Restricted to [ a , b ] range, Represent the state value function under the new and old parameters, respectively, which is used to calculate the expected cumulative discounted reward, represents the crop factor; Indicated in strategy The expectation operator under ; S4.

6. Repeat steps S4.1 to S4.5 until the strategy converges to the optimal solution or reaches the maximum number of training rounds.

10. A forest fire detection and extinguishing system, used to implement the method according to any one of claims 1 to 9, characterized in that: include: Fire identification module, fire point location and fire spread trend calculation module, fire extinguishing module, and re-ignition prevention module; The fire identification module collects fire data in real time through satellite remote sensing, remote video monitoring, and cruise drones, transmits it to the command center, generates infrared thermal imaging images, and makes fire judgments and fire grade classifications; The fire point location and fire spread trend calculation module marks potential fire points based on infrared thermal imaging images, sorts potential fire points according to temperature values, accurately determines the geographical coordinates of fire points based on multi-source geographic information data, and updates in real time according to the movement and changes of fire points in continuous external thermal imaging images, marks and locates newly appeared fire points; and performs fire spread simulation based on the Rothermel model, calculates the fire spread trend and updates the fire spread boundary; The fire extinguishing module constructs a drone dispatching model based on the Markov decision process, and pre-trains the drone dispatching model through the MAPPO multi-agent reinforcement learning algorithm; the cruise drone is equipped with fire-extinguishing bombs, and the command center performs optimized dispatching based on the trained drone dispatching model, decides the throwing order of each drone, takes the shortest time as the optimization goal, and collaboratively completes the investigation and extinguishing of all potential fire points; After extinguishing the open fire, the re-ignition prevention module continues to patrol the fire scene through the inspection drone to identify the re-ignition point or smoke point.

Images