A collaborative scheduling method for heterogeneous unmanned aerial vehicles (UAVs) in forest fire rescue based on reinforcement learning and quantum particle swarm optimization.

Abstract

This invention provides a collaborative scheduling method for heterogeneous drones in forest fire rescue based on reinforcement learning and quantum particle swarm optimization, relating to the field of collaborative scheduling. The method includes: constructing a cellular automaton fire spread model based on wind speed, vegetation, and slope, and generating a priority sequence of burning cells; establishing a three-state transition model for multiple types of drones and constructing a comprehensive cost by combining combustion loss and flight time; implementing task allocation and trajectory solving through a quantum particle swarm optimization algorithm enhanced by reinforcement learning; and employing a rolling update mechanism to correct the scheduling scheme in real time and trigger fault-tolerant replanning in abnormal situations, thereby improving the collaborative firefighting efficiency of heterogeneous drone swarms.

Description

Technical Field

[0001] This invention relates to the field of collaborative scheduling, specifically to a collaborative scheduling method for heterogeneous unmanned aerial vehicles (UAVs) in forest fire rescue based on reinforcement learning and quantum particle swarm optimization. Background Technology

[0002] Forest fire control is an important area of ​​emergency management. With the development of drone technology, using aerial platforms for fire monitoring and control is becoming a trend. The rapid response capability of this technology in complex environments has attracted attention, prompting related research to continuously expand application scenarios and collaborative models to cope with higher risks.

[0003] Existing methods mostly rely on the cooperation of ground equipment and a single aerial platform to complete fire response; they usually infer fire changes through rule models or fixed procedures and arrange operation sequences according to predetermined logic; such solutions are feasible in simple environments and are widely used, but their flexibility is limited under conditions of multi-source information coupling.

[0004] Existing solutions often struggle to handle dynamically changing environmental information and multiple task requirements simultaneously in complex scenarios; they are prone to uneven response or delayed allocation in terms of resource coordination and timing; and they also lack stability and adaptability in overall decision-making when uncertainties increase. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a heterogeneous UAV collaborative scheduling method for forest fire rescue based on reinforcement learning and quantum particle swarm optimization. This method solves the problems mentioned in the background technology, such as the difficulty in simultaneously handling dynamically changing environmental information and multiple task requirements in complex scenarios, and the tendency for uneven response or delayed allocation in resource coordination and timing.

[0006] To achieve the above objectives, this invention provides the following technical solution: a heterogeneous UAV collaborative scheduling method for forest fire rescue based on reinforcement learning and quantum particle swarm optimization, comprising the following steps:

[0007] S1. Based on the cellular automata model, the dynamic spread prediction of forest fires is carried out by comprehensively considering factors such as wind speed, vegetation density and terrain slope, and the edge burning cell priority sequence is output in real time.

[0008] S2. Based on the edge burning cell priority sequence, a heterogeneous resource pool is established for various types of UAVs with different storage capacity and flight speed, and a three-state transition model between water intake state, departure state and operation state is constructed, wherein the departure state includes the UAV flying from the water source to the target cell and returning from the target cell to the water source.

[0009] S3. Taking the cost of forest burning losses and the cost of drone flight time as dual objectives, a normalized comprehensive cost function is formed to quantitatively evaluate the comprehensive performance of different scheduling schemes.

[0010] S4. Using a quantum particle swarm optimization algorithm enhanced by reinforcement learning, under the constraints of the heterogeneous resource pool and the three-state transition model, the algorithm aims to minimize the normalized comprehensive cost function to quickly solve the firefighting task allocation and flight path timing.

[0011] S5. Based on the optimal scheduling scheme obtained from the solution, according to a fixed period... The system continuously updates the fire situation and the location of the drone swarm, recalculates the scheduling plan in a closed loop, and issues waypoint instructions in real time to continuously suppress the fire until all cells are in an unburned, burnt-out, or non-combustible state.

[0012] Furthermore, in step S1, the cellular automaton model discretizes the forest area into cellular grids, with each cell defining five states: unburned, burning, burnt out, under rescue, and non-burnable. It also simulates fire propagation based on the Moore neighborhood rule and uses state transition rules for state transition.

[0013] Furthermore, the state transition rules include:

[0014] The burning state cell ignites neighboring cells with a burning probability, which is calculated based on the base burning probability, vegetation type influence factor, vegetation density influence factor, wind speed correction factor, and slope correction factor.

[0015] Cells in a burning state that have been assigned to drone rescue are now in the "rescue in progress" state; cells in a burning state that have not been assigned to drone rescue are now in the "burned out" state.

[0016] Cells in the rescue state prevent the fire from spreading; cells in the burnt-out state cannot be rekindled; cells in the unburned state remain unburned when all their neighboring cells are also unburned; cells transitioning from the rescue state cannot be rekindled; cells in the non-flammable state remain unchanged; all the above cell states follow a fixed cycle. It will be updated on a rolling basis.

[0017] Furthermore, in step S2, the heterogeneous resource pool includes at least three types of drones, each with different storage capacity, flight speed, and quantity; the drones have three states during the rescue process: water collection state, departure state, and operation state, and the states are switched in the following order: water collection state to departure state, departure state to operation state, and operation state to water collection state.

[0018] Furthermore, the at least three types of drones include drones with different storage capacities and flight speeds; and when drones are used for firefighting, if there is remaining water, a greedy strategy is adopted to prioritize spraying the cells with the highest probability of combustion in the neighborhood of the target cell.

[0019] Furthermore, in step S3, the dual objectives include combustion loss cost and flight time cost. The combustion loss cost is quantified and accumulated based on the inherent value of each cell, and the flight time cost is calculated based on the UAV's flight path length and flight speed. After normalization, the two constitute a comprehensive optimization objective to balance rescue efficiency and resource consumption.

[0020] Furthermore, in step S4, the reinforcement learning-enhanced quantum particle swarm optimization algorithm includes a particle position update mechanism based on quantum particle swarm optimization, which introduces the average optimal position of all particles and the contraction and expansion coefficients to enhance the global search capability. It also combines the reinforcement learning module to construct a Q-table and adopts a reward mechanism with particle fitness as the state, position update amount as the action, and global fitness improvement as the reward to dynamically adjust the particle position update strategy to avoid premature convergence of the algorithm.

[0021] Furthermore, the execution flow of the reinforcement learning-enhanced quantum particle swarm optimization algorithm includes: population initialization and fitness calculation, sorting and hierarchizing by fitness to construct a Q-table, iteratively generating candidate positions, calculating rewards and updating the Q-table, selecting position update methods based on rewards, until the maximum number of iterations is reached, and outputting the optimal scheduling scheme, including UAV task allocation and paths.

[0022] Furthermore, in step S5, the rolling update cycle includes periodically running fire spread simulation and periodically running reinforcement learning-enhanced quantum particle swarm optimization algorithm scheduling; when the drone malfunctions or communication is interrupted, a fault tolerance mechanism is triggered to re-plan the water intake and fire extinguishing sequence using the remaining capacity to maintain the rescue rhythm.

[0023] Furthermore, the heterogeneous UAV forest fire rescue collaborative scheduling method based on reinforcement learning and quantum particle swarm optimization also includes priority rules for prioritizing the rescue of edge burning cells: random selection is adopted among cells of the same value to ensure fairness and coverage; and under the hybrid scheduling architecture of offline contingency plans and online correction, the location of UAVs and the evolution of the fire site are continuously perceived, and the task list is continuously refreshed to dynamically respond to changes in the fire site.

[0024] This invention provides a heterogeneous UAV collaborative scheduling method for forest fire rescue based on reinforcement learning and quantum particle swarm optimization, which has the following beneficial effects:

[0025] 1. This invention achieves dynamic fire prediction by constructing a cellular automata model that integrates wind speed, vegetation density, and slope; improves the execution efficiency of firefighting tasks by using a heterogeneous UAV three-state transition model; achieves a balance between rescue efficiency and resource consumption by using a dual-objective normalized comprehensive cost function; avoids premature convergence and improves global optimization ability by using a quantum particle swarm optimization algorithm enhanced by reinforcement learning; and achieves dynamic response and continuous fire suppression at the fire site by rolling updates and closed-loop scheduling.

[0026] 2. This invention employs cellular automata with finely divided states and multi-factor correction of fire propagation probability to improve the accuracy of fire simulation; it adopts differentiated UAV resource pools and greedy fire extinguishing strategies to improve fire extinguishing coverage efficiency; and it uses a fusion algorithm of reinforcement learning and quantum particle swarm optimization to improve the speed and accuracy of scheduling solutions. Attached Figure Description

[0027] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0028] Figure 1 This is a flowchart of the present invention;

[0029] Figure 2 A flowchart illustrating the fire propagation process of a cellular automaton;

[0030] Figure 3 Rules for mutual conversion between cell neighborhoods;

[0031] Figure 4 This is a flowchart of the drone's state transition process.

[0032] Figure 5 Box plot comparing RLQPSO with other algorithms. Detailed Implementation

[0033] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Example 1:

[0035] A collaborative scheduling method for heterogeneous unmanned aerial vehicles (UAVs) in forest fire rescue based on reinforcement learning and quantum particle swarm optimization includes the following steps:

[0036] S1. Based on the cellular automata model, the dynamic spread prediction of forest fires is carried out by comprehensively considering factors such as wind speed, vegetation density and terrain slope, and the edge burning cell priority sequence is output in real time.

[0037] S2. Based on the edge burning cell priority sequence, a heterogeneous resource pool is established for various types of UAVs with different storage capacity and flight speed, and a three-state transition model between water intake state, departure state and operation state is constructed, wherein the departure state includes the UAV flying from the water source to the target cell and returning from the target cell to the water source.

[0038] S3. Taking the cost of forest burning losses and the cost of drone flight time as dual objectives, a normalized comprehensive cost function is formed to quantitatively evaluate the comprehensive performance of different scheduling schemes.

[0039] S4. Using a quantum particle swarm optimization algorithm enhanced by reinforcement learning, under the constraints of the heterogeneous resource pool and the three-state transition model, the algorithm aims to minimize the normalized comprehensive cost function to quickly solve the firefighting task allocation and flight path timing.

[0040] S5. Based on the optimal scheduling scheme obtained from the solution, according to a fixed period... The system continuously updates the fire situation and the location of the drone swarm, recalculates the scheduling plan in a closed loop, and issues waypoint instructions in real time to continuously suppress the fire until all cells are in an unburned, burnt-out, or non-combustible state.

[0041] Preferably, the cellular automaton grid and cell state definition and operation details are as follows: In step S1, the cellular automaton model discretizes the target forest area into a 300×300 two-dimensional grid according to rules. Each grid cell is numbered by row and column coordinates and stores cell state information and environmental parameters. The cell defines five states: unburned, burning, burnt out, under rescue, and non-burnable. The neighborhood determination adopts the eight-neighbor rule of Moore's neighborhood to determine the set of adjacent cells and is used as the propagation criterion. The time step is 20 seconds to synchronously update all cell states and serve as the basic unit for rolling updates. The state transition is executed item by item according to rules R1 to R8 and implemented in the following order:

[0042] First, wind speed, vegetation density, terrain slope, and basic combustion probability are read from all cells as input parameters, and then the combustion probability is calculated. The calculation formula is:

[0043] ;

[0044] in, Based on the probability of combustion, Vegetation type influencing factors As a factor influencing vegetation density, For wind speed correction factor, This is the slope correction factor;

[0045] Subsequently, it is determined whether the cells in the burning state have been assigned to a rescue mission. If they have been assigned, the cell state is set to the rescue state and the fire spread from the cells is prevented in subsequent time steps. If they have not been assigned, the cell state is set to the burnt-out state in the next time step and marked as unreignitable. The unburned cells remain unburned when all their neighbors are unburned. The non-combustible cells remain non-combustible at any time step without changing. The condition for determining the edge burning cells is that they are in the burning state and have at least one neighboring cell in the unburned state. The priority sequence of the edge burning cells is generated by sorting all edge burning cells in descending order of burning probability. If the burning probabilities are the same, they are sorted in descending order of cell intrinsic value to generate a stable priority sequence.

[0046] Preferably, the fire propagation probability and state transition details are as follows: The cellular automaton reads four inputs at each time step: wind speed, vegetation density, terrain slope, and basic combustion probability. Each input is then normalized from minimum to maximum to map the values ​​to the 0-1 range. After normalization, the vegetation type influence factor is... Factors affecting vegetation density Multiply each factor by the corresponding normalized value to obtain the vegetation-related contribution value, and then add the wind speed correction factor. With slope correction factor The environmentally relevant contribution values ​​are obtained by multiplying each value by its corresponding normalized value; all the obtained contribution values ​​are combined using a combination of multiplication and addition to obtain the combustion impact factor, and the combustion impact factor is then compared with the basic combustion probability. Substitute the values ​​into the combustion probability function to calculate the cell ignition probability;

[0047] When the cell ignition probability is greater than or equal to the determination threshold, the cell is marked as burning. The determination threshold is 0.5 and can be set to any value between 0 and 1 in the configuration parameters. When a drone is assigned to a cell in a burning state, the cell state is updated to "rescue in progress" and the cell is excluded from any propagation determination during the rescue in progress state to prevent the fire from spreading from the cell to the neighborhood. If a cell in a burning state is not assigned by a drone at the end of the time step, the cell state is updated to "burnout" and a "cannot be re-ignited" mark is set for the cell. The "cannot be re-ignited" mark makes the cell not participate in the burning probability calculation in any subsequent time step and remains in the burnout state. A cell in an unburned state remains unburned when all eight of its neighbors are in an unburned state. A cell in a non-burnable state remains in a non-burnable state at any time step without undergoing a state transition. When multiple edge-burning cells have the same ignition probability, they are sorted in descending order of their inherent value to determine the priority sequence.

[0048] Preferably, the implementation details of the heterogeneous UAV resource pool and the three-state transition model are as follows: In step S2, the heterogeneous UAV resource pool includes three types of UAVs: UAV1, UAV2, and UAV3. UAV1 has a storage capacity of 8 units, a flight speed of 50 km / h, and a quantity of 2 units; UAV2 has a storage capacity of 9 units, a flight speed of 40 km / h, and a quantity of 3 units; and UAV3 has a storage capacity of 10 units, a flight speed of 30 km / h, and a quantity of 3 units. Each UAV maintains the variable of remaining water volume and current state identifier. The initial state is water collection state. The water collection action is performed at the water source location point, and the remaining water volume value is set to the corresponding storage capacity value, and the state is updated to departure state. The departure state is divided into departure segment and return segment. The departure segment calculates the path length from the water source location point to the target cell center point and maps it to the flight mileage using Euclidean distance. The flight time is equal to the flight mileage divided by the UAV's flight speed, and the UAV position is updated by linear interpolation within the time step. Upon reaching the target cell, the status is updated to operational status and a spraying action is executed. The spraying action first consumes the remaining water value of the target cell until the target cell's fire extinguishing flag is set. If the remaining water value is greater than zero, the remaining water is preferentially allocated to the cell with the highest combustion probability among the target cell's eight neighbors according to a greedy strategy, and spraying is repeated until the remaining water value is zero. When the remaining water value is zero, the status is updated to operational status, and the return flight path length and flight time are calculated. The drone then returns to the water source location point and the status is updated to water collection status to complete a water collection-departure-operation-return cycle. During the task allocation phase, the priority sequence of edge burning cells is used as input to allocate drones from high to low priority. During the allocation process, the shortest feasible flight segment between the drone's current position and the task target is verified to ensure that the drone can reach the target cell as planned within the departure flight segment. If the drone is in operational status and the remaining water value is greater than zero, the fire extinguishing cycle must not be forcibly interrupted, and the current fire extinguishing cycle should be completed first before participating in a new task allocation.

[0049] Preferably, the dual objectives in step S3 are combustion loss cost and UAV flight time cost; the combustion loss cost is the total combustion loss formed by summing the inherent value of each cell in the combustion state and the burnout state; the flight time cost is the flight time of a single UAV obtained by dividing the flight path length of each UAV by the corresponding flight speed, and then summing the results for all UAVs participating in the mission to form the total flight time; the total combustion loss and the total flight time are respectively normalized from minimum to maximum to map the values ​​to the interval between 0 and 1 to eliminate the difference in dimensions;

[0050] Combustion loss cost Its formula is expressed as: ;

[0051] in, For cells Its inherent value Represented as the newly generated set of combustion cells;

[0052] Flight time cost Its formula is expressed as: ;

[0053] in, This indicates selecting the maximum value in the set. Indicates the current location of the drone. This indicates the drone's location in the previous time interval. Let i represent the flight speed of the drone.

[0054] After normalization, a linear weighted synthesis is performed to generate a normalized comprehensive cost function for evaluating the scheduling scheme. The normalized comprehensive cost function is as follows:

[0055] ;

[0056] in, , For parameter factors, , These are the weighting coefficients; It is a common multiple function used to unify the relationship between flight time and the cost of burning new forests;

[0057] Particle position update

[0058] The fitness mapping formula is an improvement on the quantum particle swarm optimization algorithm, which updates particle positions. Its formula is expressed as:

[0059] ;

[0060] in, To introduce the average optimal particle position. This represents the guiding position of the i-th particle at time t. This represents the position of the i-th particle at iteration number t. These are weighting coefficients that control the step size. It is a random number;

[0061] ;

[0062] in, It is a random number. Indicates the first The optimal position of the i-th particle at iteration number t. Indicates the first The dimension is at a local optimum at iteration number t. , It is a random number.

[0063] In this embodiment, quantum behavior is used to enhance the global search capability and avoid premature convergence; fitness is defined as the complement mapping of the comprehensive cost function so that reinforcement learning can enhance the quantum particle swarm optimization algorithm to maximize the search.

[0064] Preferably, the implementation details of the reinforcement learning-enhanced quantum particle swarm optimization algorithm are as follows: In step S4, the reinforcement learning-enhanced quantum particle swarm optimization algorithm is named RLQPSO. A particle represents a feasible solution vector for UAV task allocation and flight path timing. Population initialization generates candidate solutions according to a preset number of particles, and the fitness of each particle is calculated. The fitness is evaluated using a normalized comprehensive cost function. The particle fitness is used as the reinforcement learning state input Q-table, the position update amount is used as the action input Q-table, and the global fitness improvement is used as the reward input Q-table. In each iteration, the algorithm executes in the following order: The population is sorted by particle fitness and hierarchically classified according to fitness intervals to construct Q-table entries; Candidate position update expressions are calculated based on the quantum particle swarm position update mechanism, and the average optimal position of all particles and the contraction / expansion coefficient are introduced to control the search range; After calculation, Q-table entries are updated with the reward magnitude, and position updates are performed based on the selected action output from the Q-table; After position updates, particle fitness is recalculated, and the global optimal record is updated based on fitness changes; Iteration continues until the maximum number of iterations is reached, and the optimal scheduling scheme is output according to the termination criterion.

[0065] Preferably, the rolling update cycle and fault-tolerant scheduling implementation details are as follows: Step S5, the rolling update cycle includes two parallel timed tasks and a fault handling process. The first timed task is a fire spread simulation task, which is triggered once every 20 seconds and sequentially reads the current wind speed, vegetation density, and terrain slope values ​​of all cells to calculate the combustion probability of each cell and update the cell state. The second timed task is a scheduling recalculation task, which is triggered once every 10 seconds and calls the reinforcement learning augmented quantum particle swarm optimization algorithm RLQPSO to evaluate the existing UAV allocation scheme with a comprehensive cost function and output the updated UAV task allocation and flight path sequence. The fault handling process detects a UAV fault. In the event of a communication interruption, immediately mark the lost or malfunctioning drone as unavailable and clear its task queue. Then, according to the priority sequence of edge burning cells and combined with the remaining water volume of the drone and the feasible flight time from the current position to the target cell, the remaining available drones are redistributed. The redistribution uses a greedy scheduling strategy to prioritize the allocation of drones that can arrive in the shortest flight time and have enough water to complete the initial firefighting task, while reserving subsequent resupply segments. If there are high-priority edge burning cells that cannot be covered during the redistribution process, the cells are marked as awaiting continuous attention and their priority is increased in the next scheduling recalculation. The rolling update cycle and fault handling process are executed cyclically until all cell states are returned to the unburned state, burn-out state, or non-combustible state.

[0066] Preferably, the implementation details of the hybrid scheduling of edge priority rules, offline contingency plans, and online correction are as follows: The scheduling strategy first identifies all cells in the burning state and with at least one unburned cell in their eight neighborhoods as edge burning cells. For each edge burning cell, the burning probability and cell intrinsic value are calculated to form a preliminary priority value. The preliminary priority value uses the burning probability as the primary sorting key and the cell intrinsic value as the secondary sorting key when the burning probabilities are the same to form a stable priority sequence. The priority sequence is used as the task allocation input for the scheduling module. In the case of the same priority, a pseudo-random selection function is used to randomly select target cells from the same priority cell set to ensure the fairness and coverage of task allocation. The offline contingency plan component archives multiple sets of typical fire evolution results and corresponding UAV task allocation schemes to form a contingency plan library. The entries in the contingency plan library are used as the initialization candidate solution set for RLQPSO population initialization during scheduling recalculation tasks to accelerate convergence. The online correction component reads the UAV position, remaining water volume of the UAV, and the latest cell state in real time during each scheduling recalculation task to calculate the similarity between the current situation and the entries in the contingency plan library and selects the contingency plan entry with the highest similarity as the starting solution. Subsequently, RLQPSO performs local optimization on the starting solution and outputs the final scheduling scheme. The online correction and offline contingency plan work together until all cell states are in an unburned state, a burnt-out state, or a non-combustible state.

[0067] Preferred, reinforcement learning-based enhanced quantum particle swarm optimization algorithm's rule-based execution details are as follows: A candidate solution vector with a preset number of particles is generated through population initialization, and the fitness of each candidate solution is calculated as the initial global optimum and individual optimum record; if the particle fitness is in the lowest fitness group after sorting, Q-table entries are constructed hierarchically according to fitness intervals, and particle fitness is mapped to Q-values; high-potential solutions are screened using the Q-table; discrete action sets are defined by particle position update amounts, and equal action sets are written into Q-table entries for selection; in each iteration, candidate positions are calculated using the quantum particle swarm position update expression, and the average optimal position of all particles and contraction / expansion coefficients are introduced to control the convergence range; if the global fitness improvement after position update is less than a preset minimum improvement threshold, then... The action with the highest reward in the Q-table replaces the current action to increase exploration intensity; if the global fitness improvement is greater than or equal to a preset improvement threshold, the Q-table is updated with the reward magnitude and the current action is maintained or strengthened to consolidate the local optimum; if in any iteration all particle positions do not change significantly within several consecutive iterations, a random resampling action based on the Q-table is triggered to avoid premature convergence; iteration continues until the maximum number of iterations is reached or the fitness change is lower than the termination threshold, and the UAV task allocation and trajectory sequence represented by the current global optimum are output; if the termination criterion is met, the final scheduling scheme is derived according to the optimal output rule, where the reward mechanism of reinforcement learning is defined as four elements, including: Agent is an individual particle in the particle swarm; State is the particle fitness value. Action is the direction of particle velocity update; Reward is the reward value. This reflects the extent of improvement in overall adaptability.

[0068] Preferably, the offline contingency plan library and online correction collaborative working details are as follows: Multiple typical fire evolution sequences and corresponding UAV task allocation and flight path timing schemes are pre-archived in the offline contingency plan library. If a scheduling recalculation task is triggered, several contingency plans are extracted from the library according to the contingency plan matching level during the RLQPSO population initialization phase to accelerate convergence. The online correction module reads the latest state of all cells in the current 300×300 grid, the current position of all UAVs, and the remaining water value during each scheduling recalculation task, and calculates the matching score between the current situation and the contingency plan entries. If a match is found... If the highest-scoring plan item meets the preset matching threshold, it is used as the initial solution and local optimization is performed in RLQPSO to output the final scheduling scheme. If the highest-scoring plan item does not meet the preset matching threshold, the population is initialized with random initialization and a mixture of some plans to ensure exploration diversity. In the same priority cell set, task allocation uses pseudo-random selection function sampling to ensure fairness and coverage. The pseudo-random selection function seed is generated by the current time step index to ensure reproducibility. The offline plan library entries are updated by generating the final scheme after each scheduling execution and archived by timestamp to expand the plan library content.

[0069] The RLQPSO population initialization is as follows: Let the particle swarm size be NP and the dimension be D, corresponding to the UAV scheduling decision variables, such as target cell coordinates and flight path, and the corresponding initial particle positions. and initial particle velocity The speed update formula is as follows:

[0070] ;

[0071] in For inertial weights, , For acceleration coefficient, This is the optimal position in the particle's history. The globally optimal position; The random number is a uniformly distributed random number between 0 and 1, which introduces randomness into the speed update.

[0072] The position update formula is: .

[0073] Preferably, the scheduling termination conditions, fault tolerance marking, and priority enhancement details are as follows: A rolling update mechanism is used to update cell states every 20 seconds and perform RLQPSO scheduling recalculation and issue UAV waypoint instructions every 10 seconds. If a UAV malfunctions or communication is interrupted, it is immediately marked as unavailable and its task queue is cleared. Then, tasks are redistributed according to the edge-burning cell priority sequence, combined with the current position of the remaining available UAVs and the remaining water volume, using a greedy scheduling rule. The greedy scheduling rule prioritizes UAVs with the shortest flight time to reach the target cell and sufficient water volume to meet the initial firefighting needs. If, during the redistribution process, a high-priority edge-burning cell cannot be covered by any remaining UAV in a single flight, the edge-burning cell is marked for continuous monitoring, and its priority weight is increased in the next scheduling recalculation to ensure priority allocation. When all cells within the 300×300 grid are classified as unburned, burnt out, or non-flammable, scheduling is terminated, and a final loss and flight time statistics report is output.

[0074] Example 2:

[0075] This invention provides a heterogeneous UAV collaborative scheduling method for forest fire rescue based on reinforcement learning and quantum particle swarm optimization. The scenario involves a single-point sudden fire in a mountainous forest area spreading eastward. The initial location of the fire point is at grid coordinates row 150, column 150. The initial wind speed is 10 meters per second. The initial vegetation density is represented by values ​​from 0 to 1. The initial terrain slope is 15 degrees. Water source locations are distributed at grid edge coordinates row 1, column 150 and row 300, column 150. The deployed UAV resource pool includes two UAV1s, three UAV2s, and three UAV3s. The UAV1s have a storage capacity of 8 units and a flight speed of 50 kilometers per hour. The V2 has a storage capacity of 9 units and a flight speed of 40 km / h, while the UAV3 has a storage capacity of 10 units and a flight speed of 30 km / h. The implementation steps begin with grid discretization. First, the target forest area is divided into a 300×300 two-dimensional grid, and each grid cell is initialized to an unburned state, with its inherent value and environmental parameters loaded. Then, the cell corresponding to the initial fire point is set to a burning state, and fire spread simulation and scheduling initialization are triggered synchronously at time step 0. The fire spread simulation is executed cyclically every 20 seconds. The process involves reading the wind speed, vegetation density, terrain slope, and basic combustion probability of each cell and normalizing them from minimum to maximum. The input is mapped to the interval between 0 and 1. Then, the vegetation-related contribution value and the environment-related contribution value are calculated and combined in a multiplicative-additive manner to obtain the combustion impact factor. This combustion impact factor and the basic combustion probability are substituted into the combustion probability function to obtain the cell ignition probability. Edge-burning cells are determined to be in a burning state and have at least one unburned cell in their eight neighborhoods. The edge-burning cell priority sequence is obtained by sorting all edge-burning cells from largest to smallest ignition probability, and then, if the ignition probabilities are the same, from largest to smallest cell intrinsic value. This priority sequence is used as the scheduling input to the scheduling module. Scheduling initialization at time step 0 uses several plans from the offline plan library as the initial population. Candidate solutions are combined with random initialization to generate an RLQPSO population. RLQPSO uses particle vectors to represent UAV task allocation and flight path timing, and calculates the fitness of each particle. The fitness is evaluated by a normalized comprehensive cost function. After the RLQPSO population is initialized, it enters an iterative process. The iterative process includes sorting particles by fitness and constructing a Q-table in a hierarchical manner, generating candidate positions based on the quantum particle swarm position update expression and introducing the average optimal position of all particles and contraction and expansion coefficients to control the search range, using particle fitness as the Q-table state, position update amount as the Q-table action, and global fitness improvement as the Q-table reward, and updating Q-table entries and selecting actions according to the reward.After each position update, the particle fitness is recalculated and the global optimal record is updated based on the fitness change. If the global fitness improvement is less than the preset minimum improvement threshold, the action with the greatest reward is selected in the Q table to increase the exploration intensity. If the global fitness improvement is greater than or equal to the preset improvement threshold, the Q table is updated with the reward and the current action is maintained or strengthened to consolidate the local optimum. The iteration continues until the maximum number of iterations is reached or the fitness change is lower than the termination threshold, and the optimal scheduling scheme is output according to the termination criterion.

[0076] The optimal scheduling scheme includes the water intake point, departure route, target cell sequence, and spray volume allocation for each UAV. After the task is issued, the UAV executes the water intake-departure-operation-return cycle according to the plan. The water intake action is performed at the water source location point, and the remaining water value is set to the corresponding storage capacity and the status is updated to departure state. The flight time is calculated by dividing the Euclidean distance of the departure segment by the corresponding flight speed, and the UAV position is updated by linear interpolation within the time step. After arriving at the target cell, the spraying action is performed, prioritizing the consumption of water in the target cell until the target cell's fire extinguishing flag is set. When the remaining water is greater than zero, the remaining water is allocated to the cell with the highest ignition probability in the target cell's eight neighborhoods according to a greedy strategy until the water is exhausted or there are no burning cells in the neighborhood. When the water is exhausted, the status is updated to operation and the UAV returns to the water source location point to complete the resupply before entering the next cycle. During the operation, an RLQPSO recalculation is triggered every 10 seconds to update the UAV task allocation and flight path sequence, and a fire spread simulation is performed every 20 seconds to update the cell status and generate a new edge burning cell priority sequence.

[0077] If a drone malfunction or communication interruption is detected during operation, the drone is immediately marked as unavailable and its task queue is cleared. Then, the remaining available drones are redistributed according to the priority sequence of edge burning cells, combined with the drone's current location, remaining water volume, and feasible flight time to the target cell using a greedy scheduling rule. During the redistribution process, priority is given to drones that can reach the target cell in the shortest flight time and have sufficient water volume to meet the initial fire extinguishing requirements. High-priority edge burning cells that cannot be covered in a single flight are marked as needing continuous attention so that their priority weight can be increased in the next recalculation to ensure continuous attention. The operation continues until all cells in the 300×300 grid are classified as unburned, burnt out, or non-combustible, and then terminates and outputs a final report on the combustion loss and total drone flight time.

[0078] Example 3:

[0079] This embodiment is based on Embodiment 2: the scenario involves multiple scattered fires accompanied by a sudden change in wind direction. The initial fire points are located at grid coordinates row 120, column 140 and row 180, column 160. The initial wind speed at both fire points is 8 meters per second, and at the 50th time step, the wind direction changes from east to northwest, and the wind speed is adjusted to 12 meters per second. The vegetation densities near the fire points are 0.9 and 0.75, respectively. The terrain slopes near the fire points are 20 degrees and 10 degrees, respectively. The water source location point is located at the center of the grid at row 150, column 150, with a backup water source reserved at the edge of the grid at row 1, column 150. The UAV resource pool is the same as in Embodiment 1, containing two UAV1s, three UAV2s, and three UAV3s, maintaining the same storage capacity and flight speed parameters. Implementation steps... First, at time step 0, grid discretization and cell parameter initialization are performed, and the cells corresponding to the two initial fire points are set to the burning state. The fire spread simulation runs every 20 seconds, and the input parameters are processed according to minimum to maximum normalization to calculate the ignition probability of each cell. During scheduling initialization, the offline contingency plan library contains multiple sets of typical multi-point fire evolution entries. In the population initialization stage, several contingency plans most similar to the current initial fire point distribution are extracted from the contingency plan library as the main initial candidate solutions for the population. Several randomly generated particles are mixed in the population to ensure diversity. Contingency plan matching is achieved by calculating the similarity score of the cell burning distribution and setting the matching threshold to 0.7 to control the adoption of contingency plans. If the highest matching score is lower than 0.7, the population is initialized with a mixture of contingency plans and randomization. To ensure exploration capabilities, the RLQPSO iteration process is consistent with Example 1, and the average optimal position and contraction / expansion coefficient of all particles are introduced in the position update to balance global and local searches. During the operation, after the sudden change in wind direction and speed at the 50th time step, the fire spread simulation immediately reflects the new wind speed value and recalculates the ignition probability of all cells. The priority sequence of edge burning cells is reordered according to the ignition probability and the inherent value of the cells and used as the input for the next round of scheduling. The scheduling recalculation task is triggered every 10 seconds, and after the sudden change in wind direction, special emphasis is placed on the priority allocation of newly generated edge cells with high ignition probability. The priority allocation strategy considers the flight time of the UAV from the current position to the target cell, the remaining water volume of the UAV, and the slope on the arrival path during the allocation process to ensure... Upon arrival, sufficient water is available for initial fire suppression. The spraying operation still employs a greedy strategy, prioritizing spraying the cells with the highest ignition probability in the neighborhood of the target cell when the remaining water is greater than zero. If a drone malfunctions after a sudden change in wind direction, the fault-tolerant process marks the lost drone as unavailable and clears its task queue. Subsequently, the remaining drones are redistributed using a greedy scheduling rule. During the redistribution, high-priority edge cells that cannot be covered in a single flight are marked as requiring continuous attention, and their priority weight is increased in the next round of scheduling recalculation to ensure continuous attention. During the replanning process, the plan entry in the offline plan library that is most similar to the post-mutation situation is used as population initialization to accelerate RLQPSO convergence, and the Q-table entries are updated according to the reward magnitude during the recalculation iteration.

[0080] Example 4:

[0081] This invention provides a heterogeneous UAV collaborative scheduling method for forest fire rescue based on reinforcement learning and quantum particle swarm optimization. The specific steps are as follows:

[0082] Step 1. Fire Simulation Initialization

[0083] A model based on cellular automata is constructed to facilitate subsequent drone scheduling and firefighting / rescue route planning. The specific steps are as follows:

[0084] Step 11. Mesh and Parameter Settings

[0085] Step 111. Divide the forest area into a 300×300 discrete cell grid. Each cell represents a specific geographic region, and its state is defined by numbers W1-W5, as shown in Table 1 below:

[0086] Table 1 Cell State Rules

[0087] Status number Status Name Status Description W1 Unburned state The cells were unaffected by the flames and showed no signs of burning. W2 Combustion state A cell that is currently burning will ignite surrounding cells that conform to the Moore neighborhood rule. W3 Completely burned The cell has been completely burned and cannot be ignited again (following state transition rule R5). W4 Rescue status The cell is currently in the process of a drone extinguishing a fire, at which point the flames will not spread to adjacent cells (following state transition rule R4). W5 Non-flammable state Cells cannot be ignited due to their inherent properties (such as being rock or water), and their state remains unchanged (following state transition rule R8).

[0088] Step 112. Set the initial fire source location to grid coordinates [200, 200], and set the initial fire source to a single cell ignition;

[0089] Step 113. Cell intrinsic value distribution is assigned values ​​based on vegetation type and ecological importance. High-value areas are assigned higher values, such as rare tree species distribution areas, while low-value areas are assigned lower values, such as common shrub areas.

[0090] Step 114. Environmental Parameter Settings: Wind Force Coefficient =0.045, =0.131; Base combustion probability =0.58, humidity, slope and other parameters are obtained through (Vegetation type influencing factors) (Factors affecting vegetation density) Quantification of (slope influencing factors).

[0091] Step 12. For the cell state rules established in Step 11, establish state transition rules based on the Moore neighborhood rule (a cell affects its 8 neighboring cells). The state transition is achieved through rules R1-R8, as shown in Table 2 below:

[0092] Table 2 Cell state transition rules

[0093]

[0094] Step 13. Quantification of Combustion Loss

[0095] Based on steps 112 and 113, according to the intrinsic value of the cell Establish a combustion cost function to calculate fire losses: ;

[0096] in, This represents the cell state matrix of the forest region at time t. This is represented as the rescue coverage state matrix of the UAV at time t;

[0097] Add burnout cells in this step Summation is used for subsequent algorithm optimization of the objective function.

[0098] Step 2. Configure parameters and perform state transitions for heterogeneous UAVs.

[0099] Step 21. UAV Parameter Definition

[0100] Considering that in reality, it is rare to have batches of firefighting drones with completely identical performance, three types of heterogeneous drones (UAV1-UAV3) were considered during the modeling process. The detailed parameter settings are as follows:

[0101] UAV1: Storage capacity 8, can extinguish 8 burning cells in one go, speed 50km / h, quantity 2; UAV2: Storage capacity 9, can extinguish 9 burning cells in one go, speed 40km / h, quantity 3; UAV3: Storage capacity 10, can extinguish 10 burning cells in one go, speed 30km / h, quantity 3.

[0102] Step 22. State transition mechanism

[0103] The status of drones during rescue operations is divided into three categories, and the switching logic is as follows:

[0104] State I (Water Collection State): After the drone is filled with water from the water source, it flies towards the target cell to burn water. At this time, Load = storage capacity.

[0105] State II (Departure State): The UAV arrives at the target cell and performs firefighting operations. Load decreases as the firefighting process progresses.

[0106] State III (Operational State): After firefighting is completed, the drone returns empty to the water source to replenish water; at this time, Load=0. State I can transition to State II, State II can transition to State III, and State III can transition to State I.

[0107] Step 23. Fire Extinguishing Rules

[0108] Prioritize rescuing high-value cells: In cells on the edge of combustion, i.e., W2 state cells adjacent to unburned cells, according to... Cells with higher value are prioritized for rescue missions.

[0109] Remaining water handling: If there is still water remaining after the drone extinguishes the fire, a greedy strategy is adopted to prioritize spraying the cells with the highest probability of combustion in the target cell's neighborhood, thereby reducing water waste.

[0110] Step 3. Execution of the RLQPSO algorithm

[0111] To achieve collaborative scheduling of heterogeneous UAVs for forest fire rescue based on reinforcement learning (RL) and quantum particle swarm optimization (QPSO), this embodiment proposes a QPSO algorithm based on reinforcement learning. Building upon the basic QPSO algorithm, a reward mechanism from reinforcement learning is added, strongly rewarding solutions carrying better information, thereby achieving better convergence. Details are as follows:

[0112] Step 31. Basic Initialization of Particle Swarm Optimization Algorithm

[0113] Let the particle swarm size be NP and the dimension be D (corresponding to the UAV scheduling decision variables, such as target cell coordinates and flight path), and the corresponding initial particle positions. and speed The speed update formula is as follows:

[0114] ;

[0115] in For inertial weights, , For acceleration coefficient, This is the optimal position in the particle's history. The globally optimal position; The random number is a uniformly distributed random number between 0 and 1, which introduces randomness into the speed update.

[0116] The position update formula is: .

[0117] Step 32. Improvement of Quantum Particle Swarm Optimization Algorithm

[0118] Traditional particle swarm optimization (PSO) algorithms have advantages such as fast convergence speed, but they are also susceptible to perturbations and getting trapped in local optima. By removing the velocity variable and updating particle positions through quantum behavior:

[0119] ;

[0120] in, To introduce the average optimal particle position. This represents the guiding position of the i-th particle at time t. This represents the position of the i-th particle at iteration number t. These are weighting coefficients that control the step size. A random number in the range [0, 1);

[0121] ;

[0122] in, It is a random number. Indicates the first The optimal position of the i-th particle at iteration number t. Indicates the first The dimension is at a local optimum position at iteration number t.

[0123] By sampling particle positions multiple times using the Monte Carlo method, the position with the best fitness is selected as the next generation particle position, thereby enhancing the global search capability.

[0124] Step 33. Improvements to QPSO using Reinforcement Learning (RL)

[0125] State definition: based on the fitness value of the particle swarm. As an environmental state;

[0126] Action definition: The action of the agent is to update the velocity vector based on the particle's position.

[0127] Reward mechanism: Through Calculate rewards, Let r be the global fire intensity of the forest region at time t, where r>0 indicates the update is valid and r=0 indicates it is invalid.

[0128] Q table update: , For learning rate, This is a discount factor used to guide particles to search for high-reward regions;

[0129] Position update strategy: If r > 0, retain the current position. If r=0, then through renew, , The position of the particle with the maximum Q value. The optimal particle position for fitness.

[0130] Step 34. Complete RLQPSO Process

[0131] Input parameters: population size NP, dimension D, maximum number of iterations Maxgeneration, shrinkage coefficient α, learning rate β, etc.; initialize the population and calculate fitness; sort and rank by fitness, and construct the Q-table; the iterative optimization is: generate candidate positions according to the QPSO formula. Calculate rewards Update the Q table; select the position update method based on the r value; repeat until the maximum number of iterations is reached, and output the optimal scheduling scheme (UAV task allocation and path).

[0132] Step 4. Real-time decision-making strategy

[0133] Step 41. Time Synchronization Mechanism

[0134] A fire spread simulation is run every Ts = 20 seconds to update the cell states (W1-W5) and ;

[0135] The RLQPSO scheduling algorithm runs every Ta = 10 seconds to update the UAV's position and mission status (states I-III).

[0136] Step 42. Dynamically adjust the logic

[0137] Based on the latest fire status (such as the location of newly added burning cells and the distribution of high-value areas), the RLQPSO algorithm is re-executed to generate new drone scheduling paths. If the drones have remaining water during the firefighting process, a greedy strategy is triggered to prioritize spraying the cells with the highest burning probability in the neighborhood of the target cells to suppress the spread of the fire. When the drones complete the current task state III, the task is immediately reassigned to state I to ensure efficient use of resources. The termination condition is when all burning cells have turned to burnout or the rescue is completed, and no new burning cells are generated, the simulation is terminated.

[0138] Example 5:

[0139] This embodiment provides a heterogeneous UAV collaborative scheduling method for forest fire rescue based on reinforcement learning and quantum particle swarm optimization. The effectiveness and feasibility of the invention are further illustrated by the following experiments. All experiments are implemented using MATLAB R2024a programming, running on Windows 11 Professional Edition, with hardware configuration of Intel i7-13900H 3.8GHz, 64GB RAM, and RTX 4060 graphics card. The core algorithm code has been encapsulated as a scheduling middleware, which can be used plug-and-play on edge computing nodes.

[0140] (1) Experimental Design: To verify the applicability of this invention in a real early-stage fire environment, a 300×300 cell forest grid was constructed. The fire information update cycle was 20s, and the UAV pose refresh cycle was 10s. The ignition point was uniformly set at coordinates [200,200]. Parameters such as wind speed, vegetation density, and slope were randomly selected based on measured data to form a continuously changing dynamic fire field. The UAV swarm adopted three heterogeneous models: UAV1 with a water storage capacity of 8L and a speed of 50km / h; UAV2 with a water storage capacity of 9L and a speed of 40km / h; and UAV3 with a water storage capacity of 10L and a speed of 30km / h. A total of 8 sorties were conducted, and the ratio of the UAV swarm to the fire area was consistent with the actual situation of early-stage rescue in mountainous areas.

[0141] (2) Optimization Solution: The RLQPSO scheduling core of this invention and the comparison algorithm were placed on the same simulation platform with a population size of 30 and a maximum iteration of 1000 generations. To reduce random errors, each instance was run independently 30 times, and the optimal value, average value, and variance were statistically analyzed. Table 3 shows that the optimal objective function value of RLQPSO is 5.61 × 10⁻⁶. -5 Average 3.63×10 -3 Variance 1.27×10 -5 All of them are significantly better than the comparison algorithms, and have the smallest variance, indicating that their solutions have high quality and good stability; for example Figure 5 As shown, the comparison box plots further confirm that the RLQPSO algorithm has the lowest median, the data distribution is concentrated, and there is no abnormal divergence.

[0142] Table 3 Experimental results of different algorithms

[0143] Maximum value Minimum value average value variance QPSO 3.02E-02 6.79E-03 1.40E-02 1.03E-04 PSO 2.28E-02 6.28E-04 9.29E-03 5.77E-05 MPSO 2.25E-02 6.65E-03 1.10E-02 4.93E-05 SCA 2.67E-02 1.78E-04 1.15E-02 8.68E-05 RLQPSO 1.16E-02 5.61E-05 3.63E-03 1.27E-05

[0144] (3) Combustion Cost Assessment: Fire loss was quantified using two indicators: RVB (Rescued Units / Combustion Units) and BVA (Combustion Units / Total Units). Table 4 shows the statistical mean ± standard deviation of 30 tests: RLQPSO had the highest RVB at 20.53% ± 0.31%, while its BVA was the lowest at only 4.54‰ ± 0.62‰. The Mann-Whitney U test p < 0.05, demonstrating that RLQPSO significantly outperformed the suboptimal algorithm SCA (RVB 19.65%, BVA 9.39‰) in both reducing fire loss and improving rescue efficiency.

[0145] Table 4 Comparison of Combustion Value Indicators of Different Algorithms

[0146] RVB BVA QPSO 16.33%±0.42% 12.65‰±0.53‰ PSO 8.65%±0.73% 21.02‰±0.96‰ MPSO 14.39%±0.44% 18.71‰±0.30‰ SCA 19.65%±1.07% 9.39‰±0.25‰ RLQPSO 20.53%±0.31% 4.54‰±0.62‰

[0147] (4) Capacity Impact Test of Heterogeneous UAVs: The core of the fixed scheduling algorithm is RLQPSO, with only the water storage capacity of a single UAV changed, as shown in Table 5, forming three comparison groups: Case 1 (10 units), Case 2 (8-10 units mixed with vegetation type influence factors), and Case 3 (5-7 units mixed). Table 6 shows the comparison of the running time of different combination rescue simulation operations: Case 1 only required 8 iterations to complete fire extinguishing, with a BVA of 4.30‰; Case 3 still failed to completely extinguish the fire after 252 iterations, with the BVA rising to 19.01‰. The experiment proves that under the same number of sorties and speed conditions, increasing the water storage capacity can shorten the rescue rounds and significantly suppress the spread of the fire front, verifying the rationality of the "capacity priority" principle of this invention.

[0148] Table 5 Capacity Impact Test of Heterogeneous UAVs

[0149]

[0150] Table 6 Comparison of running times for different combinations of rescue simulation operations

[0151] Case Number of runs at the end RVB BVA 1 8 15.76%±0.58% 4.30‰±1.34‰ 2 52 20.53%±0.98% 4.54‰±2.12‰ 3 252 1.48%±0.77% 19.01%±0.53%

[0152] In summary, under typical early-stage forest fire scenarios, this invention can provide a stable and low-cost heterogeneous UAV collaborative scheduling scheme within a 20-second rolling cycle, reducing the burning area by 7.3%, lowering the burning loss ratio to 4.54‰, and increasing the single-shot fire extinguishing efficiency of high-capacity UAVs by 50%, fully demonstrating the effectiveness, feasibility, and engineering promotion value of the proposed method.

[0153] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A heterogeneous unmanned aerial vehicle forest fire rescue collaborative scheduling method based on reinforcement learning and quantum particle swarm optimization, characterized in that, The method comprises the following steps: S1, based on the cellular automaton model, the factors of wind speed, vegetation density and terrain slope are integrated to predict the dynamic spread of forest fire, and the edge burning cell priority sequence is output in real time; S2, based on the edge burning cell priority sequence, a heterogeneous resource pool is established for multiple types of unmanned aerial vehicles with differentiated storage capacity and flight speed, and a three-state transition model between water taking state, departure state and operation state is constructed, wherein the departure state includes the flight of unmanned aerial vehicles from the water source to the target cell and the return of unmanned aerial vehicles from the target cell to the water source; S3, taking the forest burning loss cost and the unmanned aerial vehicle flight time cost as double objectives, a normalized comprehensive cost function is formed to quantitatively evaluate the comprehensive performance of different scheduling schemes; the double objectives include the burning loss cost and the flight time cost, the burning loss cost is quantitatively accumulated based on the inherent value of each cell, and the flight time cost is calculated based on the flight path length and flight speed of the unmanned aerial vehicle; after normalization, the two form a comprehensive optimization objective to balance the rescue efficiency and resource consumption; S4, a quantum particle swarm optimization algorithm enhanced by reinforcement learning is used to minimize the normalized comprehensive cost function as the objective under the constraints of the heterogeneous resource pool and the three-state transition model to quickly solve the fire extinguishing task allocation and flight path timing; The quantum particle swarm optimization algorithm enhanced by reinforcement learning includes a particle position updating mechanism based on quantum particle swarm optimization, the average optimal position of all particles and the contraction and expansion coefficient are introduced to enhance the global search ability, and a Q table is constructed by combining a reinforcement learning module; a reward mechanism is adopted, in which the particle fitness is used as the state, the position updating amount is used as the action, and the global fitness improvement amplitude is used as the reward, to dynamically adjust the particle position updating strategy to avoid premature convergence of the algorithm; S5、Based on the optimal scheduling scheme obtained, the fixed period is updated Rolling update fire state and UAV group position, closed loop recalculation of scheduling scheme and real-time issuance of waypoint instructions, continuous suppression of fire until all cells turn into unburned state, burned out state or non-flammable state.

2. The heterogeneous unmanned aerial vehicle forest fire rescue collaborative scheduling method based on reinforcement learning and quantum particle swarm optimization according to claim 1, characterized in that: In step S1, the cellular automaton model discretizes the forest area into a cell grid, defines five types of states for each cell: unburned state, burning state, burned-out state, rescue-in-progress state and non-combustible state, simulates fire propagation based on Moore neighborhood rules, and performs state transition using state transition rules.

3. The heterogeneous unmanned aerial vehicle forest fire rescue collaborative scheduling method based on reinforcement learning and quantum particle swarm optimization according to claim 2, characterized in that: The state transition rules include: The burning state cell ignites the neighbor cell with a burning probability, which is calculated based on the basic burning probability, the vegetation type influence factor, the vegetation density influence factor, the wind speed correction factor and the slope correction factor; the burning state cell arranged for unmanned aerial vehicle rescue is converted to the rescue-in-progress state; the burning state cell not arranged for unmanned aerial vehicle rescue is converted to the burned-out state; Rescue state cells prevent fire spread; Burnout state cells cannot be re-ignited; Unburnt state cells remain unburnt if all their neighbor cells are also unburnt; Cells transitioning from rescue state cannot be re-ignited; Non-flammable state cells remain unchanged; All the above cell states are updated at fixed intervals Rolling updates are performed.

4. The heterogeneous unmanned aerial vehicle forest fire rescue collaborative scheduling method based on reinforcement learning and quantum particle swarm optimization according to claim 1, characterized in that: In step S2, the heterogeneous resource pool includes at least three types of unmanned aerial vehicles with differentiated storage capacity, flight speed and quantity; the unmanned aerial vehicles have three states in the rescue process: water taking state, departure state and operation state, and the states are converted in the order of water taking state to departure state, departure state to operation state, and operation state to water taking state.

5. The heterogeneous unmanned aerial vehicle forest fire rescue collaborative scheduling method based on reinforcement learning and quantum particle swarm optimization according to claim 4, characterized in that: The at least three types of unmanned aerial vehicles include unmanned aerial vehicles with differentiated storage capacity and flight speed; and when the unmanned aerial vehicles are extinguishing the fire, if there is remaining water, a greedy strategy is adopted to preferentially spray the cell with the highest burning probability in the neighbor of the target cell.

6. The heterogeneous unmanned aerial vehicle forest fire rescue collaborative scheduling method based on reinforcement learning and quantum particle swarm optimization according to claim 1, characterized in that: The execution flow of the quantum particle swarm optimization algorithm enhanced by reinforcement learning comprises population initialization and fitness calculation, fitness sorting and grading to build a Q table, iterative generation of a candidate position, reward calculation and Q table updating, selection of a position update mode according to the reward, output of an optimal scheduling scheme, including unmanned aerial vehicle task allocation and path, until a maximum number of iterations is reached.

7. The heterogeneous unmanned aerial vehicle forest fire rescue collaborative scheduling method based on reinforcement learning and quantum particle swarm optimization according to claim 1, characterized in that: In step S5, the rolling update cycle comprises periodic running of the fire spread simulation and periodic running of the quantum particle swarm optimization algorithm enhanced by reinforcement learning scheduling; when the unmanned aerial vehicle fails or communication is interrupted, a fault-tolerant mechanism is triggered, and the remaining operating capacity is used to re-plan the water taking and fire extinguishing sequence to maintain the rescue rhythm.

8. The heterogeneous unmanned aerial vehicle forest fire rescue collaborative scheduling method based on reinforcement learning and quantum particle swarm optimization according to any one of claims 1 to 7, characterized in that, The priority rule of preferentially rescuing the edge burning cell is also included: random selection is adopted between cells of the same value to ensure fairness and coverage; and under the mixed scheduling architecture of offline preplan and online correction, the unmanned aerial vehicle position and the fire field evolution are continuously perceived, the task list is rolling refreshed to dynamically respond to the fire field change.

Images