Disaster reduction rescue path planning method and system based on internet of things

Abstract

The application discloses a disaster reduction and rescue path planning method and system based on Internet of Things, and the method comprises environment interaction, network updating, reward function design, inner layer network training, outer layer optimization search and disaster reduction and rescue path planning. The application belongs to the field of path planning, and particularly relates to a disaster reduction and rescue path planning method and system based on Internet of Things. The scheme reduces instruction conflict, overshoot and oscillation by executing control instructions in sections and delaying the issuance of the control instructions. The safety and reliability of the disaster reduction and rescue path planning are improved by introducing dynamic temporary priorities, combining distance progress, heading punishment and energy perception to design sectional rewards. The unique adaptive probability adjustment of crossover and mutation operations is combined with dynamic adjustment of performance index fluctuation to adapt to complex disaster conditions on site and avoid parameters from falling into local optimum. The gene reset mechanism of the enhanced stage is introduced to continuously explore potential optimal solutions and improve path efficiency and rescue effect.

Description

Technical Field

[0001] This invention relates to the field of route planning technology, specifically to a disaster reduction and relief route planning method and system based on the Internet of Things. Background Technology

[0002] Disaster relief path planning is an intelligent decision-making process for disaster areas. Essentially, it automatically generates safe and efficient search and rescue flight paths for drones within a closed loop of real-time perception, adaptive learning, and optimized deployment. However, general disaster relief path planning methods suffer from problems such as unstable command issuance and execution control leading to control disorder, and a lack of dynamic prioritization resulting in poor accuracy. Furthermore, these methods often suffer from low parameter tuning efficiency and poor adaptability to complex, dynamic, and multi-objective disaster relief environments, ultimately leading to poor path planning effectiveness. Summary of the Invention

[0003] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides an IoT-based disaster relief path planning method and system. Addressing the problems of unstable command issuance and execution control leading to control disorder, and the lack of dynamic priorities resulting in poor accuracy in disaster relief using conventional disaster relief path planning methods, this solution reduces command conflicts and minimizes overshoot and oscillations by segmenting and delaying the issuance of control commands. It improves the safety and reliability of disaster relief path planning by introducing dynamic temporary priorities and combining distance progress, heading penalties, and energy perception to design segmented rewards. Furthermore, addressing the problems of low parameter tuning efficiency and poor adaptability to complex, dynamic, and multi-objective disaster relief environments in conventional disaster relief path planning methods, leading to poor path planning results, this solution uses unique crossover and mutation operations for adaptive probability adjustment, combined with dynamic adjustments based on performance index fluctuations, to adapt to complex disaster situation changes on-site and avoid parameters getting trapped in local optima. Finally, it introduces a gene reset mechanism in the enhancement phase to continuously explore potential optimal solutions under the instability of the rescue site environment, thereby improving path efficiency and rescue effectiveness.

[0004] The technical solution adopted by this invention is as follows: The disaster mitigation and rescue route planning method based on the Internet of Things provided by this invention includes the following steps:

[0005] Step S1: Environmental interaction;

[0006] Step S2: Network update;

[0007] Step S3: Design the reward function;

[0008] Step S4: Inner network training;

[0009] Step S5: Outer layer optimization search;

[0010] Step S6: Disaster mitigation and relief route planning.

[0011] Further, in step S1, the environmental interaction involves using IoT sensors and drones deployed in the disaster area to jointly construct a real-time environmental grid map; decomposing the action signals output by the path planning strategy network into turning, forward movement, and ascent / descent; firstly, a turning command is issued to the drone from the cloud computing node; and then a delay is applied. Then issue forward and vertical commands; throughout the entire action, at a frequency Acquire data from IoT sensor networks and UAV status for collision warning; UAV status includes position, speed, and obstacle distance; at low frequency. Record the environment state, strategy actions, and execution results as a complete interaction experience and store it in the cloud experience playback cache.

[0012] Further, in step S2, the network update involves storing accumulated experience in the replay cache of cloud nodes, randomly sampling it in batches, and using it for updating the Q network and policy network; a delayed cyclic step strategy is adopted, with the Q network and policy network updating at different frequencies; two sets of Q networks are deployed in the cloud, including an online network and a target network, with the online network updating in real time and the target network updating every [time missing]. The parameters are synchronized once per iteration; each Q-network update minimizes the deviation by using the smaller of the two outputs.

[0013] Furthermore, in step S3, the reward function design is based on the search and rescue mission, refining the reward function into a multi-objective combination; and introducing a temporary priority index. The reward function is defined piecewise as follows: ; ; ;in, It is a communication link quality indicator; It represents the remaining battery percentage; It is the Euclidean distance between the current point and the target point; It is the distance from the target point in one step; It is the horizontal angle between the drone's current heading and the target direction; It is a communication weighting coefficient; It is the energy weighting coefficient; It is a positive reward for reaching the target area; It is a negative reward for collisions; , and These refer to reaching the target area, collision, and other situations, respectively. It is the distance progress reward coefficient; d is the heading deviation penalty coefficient; d is the relative distance-to-schedule ratio. It represents the remaining target points; It is the maximum number of target points; It is the task threshold; sign(·) is the sign function.

[0014] Further, in step S4, the inner network training involves initializing the cloud computing nodes and cloud server, loading the Q network and policy network parameters; initializing the replay cache, setting the hyperparameters of learning rate, discount factor and update frequency; preparing the UAV and IoT sensor network, and starting to collect real-time environmental data; repeating the iteration until convergence or reaching the maximum number of rounds, and the network training is completed; the iteration content is: (1) the current policy network is based on the state Give an action , will the action Break it down into steering, forward, and elevation / reclining actions, with delay. The segmentation strategy is distributed to the drone; by frequency Monitor the drone's status and surrounding environment, and record collision warnings in a timely manner; using frequency (1) Record an interaction sample to the replay buffer; (2) If the Q-network or policy network update condition is met, randomly sample a batch of samples from the replay buffer; update the target using the minimum value of the dual Q-network and calculate the target value. And perform gradient descent on the online Q network parameters; use the updated Q network to guide the gradient ascent of the policy network and update the policy network parameters; after the synchronization condition is reached, copy the online Q network parameters to the target network; (3) continue to execute the next time step until the end of the whole round; at this time, the single inner layer training ends, and the performance index under this hyperparameter configuration is output; performance index Represented as: ;in, , , and These are average cumulative reward, collision rate, communication link stability, and energy consumption. , , and These are the corresponding weights.

[0015] Furthermore, in step S5, the outer layer optimization search specifically includes the following steps:

[0016] Step S51: The learning rate, discount factor, update frequency, and reward weight are set as individual chromosome parameters, and optimization evaluation is performed in parallel on the cloud. For each generation of optimized individuals, the inner network is trained in parallel on the cloud to obtain the fitness of each individual in the population. M individuals are initialized and mapped to a binary population. Performance metrics are used as individual fitness. Crossover and mutation are performed on individuals based on adaptive probabilities. For the nth generation, the crossover and mutation probabilities of individuals are expressed as: ;

[0017] ;in, It is the probability that an individual performs a crossover operation; and These are the upper and lower bounds of the crossover probability, respectively. It is the probability that an individual will perform a mutation operation; and These are the upper and lower bounds of the mutation operation, respectively; It is the average fitness value of all individuals in the current generation; It is the maximum fitness value of all individuals in the current generation; It is the maximum fitness among the two parents involved in the crossover; It is the individual's own adaptability; It is the maximum iteration algebra; It is an exponential decay factor;

[0018] Step S52: Reset enhancements; every m generations, select the middle-ranked fitness groups. Individual genes are randomly reset, and the crossover / mutation probability is temporarily increased, as shown below: ; ;in, and These are the crossover and mutation probabilities after the temporary boost to individuals during the reset enhancement phase, respectively; w is the window center. It is the shape factor;

[0019] Step S53: Set a threshold and end when the maximum number of generations is reached or the fitness fluctuation is less than the threshold; obtain the hyperparameters of the optimal individual representation; send the hyperparameters of the optimal individual representation to the cloud node, replace the learning rate, discount factor, update frequency and reward weight settings, and train the Q network and policy network.

[0020] Furthermore, in step S6, the disaster relief path planning involves deploying a pre-trained Q network and policy network at the disaster site; based on the real-time status, the policy network is directly invoked to output the optimal action; similarly, a segmented execution mechanism and IoT sensing information are used to maintain stable control and collision warning for the drone.

[0021] The disaster reduction and relief path planning system based on the Internet of Things provided by this invention includes an environment interaction module, a network update module, a reward function design module, an inner network training module, an outer optimization search module, and a disaster reduction and relief path planning module.

[0022] The environmental interaction module uses IoT sensors and drones to collaboratively construct a real-time environmental grid map and issue actions in segments to achieve environmental interaction;

[0023] The network update module uses Q network and policy network at different frequencies, combined with a dual-Q structure and replay buffer, to asynchronously update network parameters;

[0024] The reward function design module designs a reward function based on multiple objectives related to rescue needs;

[0025] The inner network training module collects environmental interaction data and continuously iterates the training network parameters to obtain performance indicators.

[0026] The outer-layer optimization search module optimizes hyperparameters, employs adaptive crossover and mutation mechanisms, and uses phased enhancement to filter for the optimal configuration.

[0027] The disaster relief and rescue route planning module deploys the optimal strategy to the disaster area and executes the rescue route planning.

[0028] The beneficial effects achieved by the present invention using the above solution are as follows:

[0029] (1) In view of the problems that the general disaster relief path planning method has not been stable enough in the issuance and execution of instructions, resulting in control disorder, and lack of dynamic priority, resulting in poor accuracy of disaster relief, this scheme reduces instruction conflict and reduces overshoot and oscillation by segmenting the execution and delaying the issuance of control instructions; by introducing dynamic temporary priority and combining distance progress, heading penalty and energy perception to design segmented rewards, the safety and reliability of disaster relief path planning are improved.

[0030] (2) In view of the problems that general disaster relief and rescue path planning methods have low parameter adjustment efficiency and poor adaptability to complex, dynamic and multi-objective disaster relief and rescue environments, resulting in poor disaster relief and rescue path planning effect, this scheme adopts a unique crossover and mutation operation to adaptive probability adjustment, combined with performance index fluctuations to dynamically adjust, adapt to complex disaster situation changes on site, and avoid parameters from getting stuck in local optima; introduces a gene reset mechanism in the enhancement stage, continuously explores potential optimal solutions under the instability of the rescue site environment, and improves path efficiency and rescue effect. Attached Figure Description

[0031] Figure 1 A flowchart illustrating the Internet of Things-based disaster mitigation and relief route planning method provided by this invention;

[0032] Figure 2 This is a schematic diagram of the Internet of Things-based disaster reduction and rescue route planning system provided by the present invention.

[0033] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0034] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0035] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0036] Example 1, see Figure 1 The present invention provides a disaster mitigation and relief route planning method based on the Internet of Things, which includes the following steps:

[0037] Step S1: Environmental Interaction; Real-time environmental grid map is constructed through collaboration between IoT sensors and drones, and actions are sent out in segments to achieve environmental interaction;

[0038] Step S2: Network update; Using Q network and policy network at different frequencies, combined with double-Q structure and replay buffer, the network parameters are updated asynchronously;

[0039] Step S3: Reward function design; Design a reward function based on multiple objectives of rescue needs;

[0040] Step S4: Inner network training; collect environmental interaction data and continuously iterate the training of network parameters to obtain performance indicators;

[0041] Step S5: Outer layer optimization search; by optimizing hyperparameters, an adaptive crossover and mutation mechanism is adopted, as well as a phased enhancement to select the optimal configuration;

[0042] Step S6: Disaster mitigation and relief route planning; Deploy the optimal strategy to the disaster area and execute the relief route planning.

[0043] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the environmental interaction utilizes IoT sensors and drones deployed in the disaster area to jointly construct a real-time environmental grid map; the action signals output by the path planning strategy network are decomposed into turning, forward movement, and ascent / descend; firstly, a turning command is issued to the drone from the cloud computing node; delay... Then, forward and vertical commands are issued to ensure the controller has sufficient time to complete pitch and turn; throughout the entire action execution, at a frequency Acquire data from IoT sensor networks and UAV status for collision warning; UAV status includes position, speed, and obstacle distance; at low frequency. Record the environmental state, strategy actions, and execution results as a complete interaction experience and store it in the cloud experience playback cache; avoid control disorder caused by frequent switching of instructions for the drone through a segmented execution mechanism to ensure the smooth movement of the rescue platform.

[0044] Example 3, see Figure 1 This embodiment is based on the above embodiment. In step S2, the network update involves storing accumulated experience in the replay cache of the cloud node, randomly sampling it in batches, and using it for updating the Q network and policy network. A delayed cyclic step strategy is adopted, with the Q network and policy network updated at different frequencies to accommodate unstable communication and computing resource fluctuations in the field. Two Q networks are deployed in the cloud, including an online network and a target network. The online network is updated in real time, and the target network is updated every [time period missing]. Each iteration synchronizes the parameters once; each Q-network update minimizes the deviation by using the smaller of the two outputs, reducing the risk of Q-value overestimation; the objective function and Q-network loss are expressed as: ; ;in, It is the target value at time t; t is time. It is a discount factor; and These are the states at time t+1 and time t, respectively; and These are the actions at time t+1 and time t, respectively. These are the target network parameters; It is the Q-value estimate of the target network for the next state; It is the loss value; These are the Q-network parameters; the Q-network parameter update is represented as: Policy network updates are represented as: The Min-Q strategy is represented as: ;in, It is the Q network learning rate; It is the learning rate of the policy network; and These are the updated parameters for the Q network and the policy network, respectively. and These are the parameters of the Q network and the policy network before the update, respectively. and They are about and The gradient; and These are the outputs of two parallel Q-networks; s and a represent the state and action, respectively.

[0045] Example 4, see Figure 1 This embodiment is based on the above embodiment. In step S3, the reward function design is based on the search and rescue mission, and the reward function is refined into a multi-objective combination. For rescue path planning, the current target point is used as the starting point and the next target point is used as the ending point. In order to improve the hit rate during critical periods, a temporary priority indicator is introduced. The higher the remaining task points and battery level, the higher the reward. The reward function is defined piecewise as follows: ; ; ;in, It is a communication link quality indicator; It represents the remaining battery percentage; It is the Euclidean distance between the current point and the target point; It is the distance from the target point in one step; It is the horizontal angle between the drone's current heading and the target direction; It is a communication weighting coefficient; It is the energy weighting coefficient; It is a positive reward for reaching the target area; It is a negative reward for collisions; , and These refer to reaching the target area, collision, and other situations, respectively. It is the distance progress reward coefficient; d is the heading deviation penalty coefficient; d is the relative distance-to-schedule ratio. It represents the remaining target points; It is the maximum number of target points; It is the task threshold; sign(·) is the sign function.

[0046] Example 5, see Figure 1 This embodiment is based on the above embodiment. In step S4, the inner network training is to initialize the cloud computing node and cloud server, load the Q network and policy network parameters; initialize the replay cache, set the hyperparameters of learning rate, discount factor and update frequency; prepare the UAV and IoT sensor network, and start collecting real-time environmental data; repeat the iteration until convergence or the maximum number of rounds is reached, and the network training is completed. The iteration content is: (1) The current policy network is based on the state Give an action , will the action Break it down into steering, forward, and elevation / reclining actions, with delay. The segmentation strategy is distributed to the drone; by frequency Monitor the drone's status and surrounding environment, and record collision warnings in a timely manner; using frequency (1) Record an interaction sample to the replay buffer; (2) If the Q-network or policy network update condition is met, randomly sample a batch of samples from the replay buffer; update the target using the minimum value of the dual Q-network and calculate the target value. And perform gradient descent on the online Q network parameters; use the updated Q network to guide the gradient ascent of the policy network and update the policy network parameters; after the synchronization condition is reached, copy the online Q network parameters to the target network; (3) continue to execute the next time step until the end of the whole round; at this time, the single inner layer training ends, and the performance index under this hyperparameter configuration is output; performance index Represented as: ;in, , , and These are average cumulative reward, collision rate, communication link stability, and energy consumption. , , and These are the corresponding weights.

[0047] By performing the above operations, this solution addresses the problems of inconsistent command issuance and execution control leading to control disorder, and poor accuracy in disaster relief due to the lack of dynamic priorities in general disaster relief path planning methods. This solution reduces command conflicts and minimizes overshoot and oscillations by segmenting and delaying the issuance of control commands. Furthermore, by introducing dynamic temporary priorities and combining them with distance progress, heading penalties, and energy perception to design segmented rewards, the solution improves the safety and reliability of disaster relief path planning.

[0048] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S5, the outer layer optimization search specifically includes the following steps:

[0049] Step S51: The learning rate, discount factor, update frequency, and reward weight are set as individual chromosome parameters, and optimization evaluation is performed in parallel on the cloud. For each generation of optimized individuals, the inner network is trained in parallel on the cloud to obtain the fitness of each individual in the population. M individuals are initialized and mapped to a binary population. Performance metrics are used as individual fitness. Crossover and mutation are performed on individuals based on adaptive probabilities. For the nth generation, the crossover and mutation probabilities of individuals are expressed as: ;

[0050] ;in, It is the probability that an individual performs a crossover operation; and These are the upper and lower bounds of the crossover probability, respectively. It is the probability that an individual will perform a mutation operation; and These are the upper and lower bounds of the mutation operation, respectively; It is the average fitness value of all individuals in the current generation; It is the maximum fitness value of all individuals in the current generation; It is the maximum fitness among the two parents involved in the crossover; It is the individual's own adaptability; It is the maximum iteration algebra; It is an exponential decay factor;

[0051] Step S52: Reset enhancements; every m generations, select the middle-ranked fitness groups. Individual genes are randomly reset, and the crossover / mutation probability is temporarily increased, as shown below: ; ;in, and These are the crossover and mutation probabilities after the temporary boost to individuals during the reset enhancement phase, respectively; w is the window center. It is the shape factor;

[0052] Step S53: Set a threshold and end when the maximum number of generations is reached or the fitness fluctuation is less than the threshold; obtain the hyperparameters of the optimal individual representation; send the hyperparameters of the optimal individual representation to the cloud node, replace the learning rate, discount factor, update frequency and reward weight settings, and train the Q network and policy network.

[0053] Example 7, see Figure 1 This embodiment is based on the above embodiment. In step S6, disaster relief path planning involves deploying a pre-trained Q network and policy network at the disaster site. Based on the real-time status, the policy network is directly invoked to output the optimal action. Similarly, a segmented execution mechanism and IoT sensing information are used to maintain stable control and collision warning for the drone.

[0054] By performing the above operations, this solution addresses the problems of low parameter tuning efficiency and poor adaptability to complex, dynamic, and multi-objective disaster relief environments in general disaster relief path planning methods, which ultimately lead to poor path planning results. It employs unique crossover and mutation operations for adaptive probability adjustment, combined with dynamic adjustments based on performance index fluctuations, to adapt to complex disaster changes on-site and avoid parameters getting trapped in local optima. Furthermore, it introduces a gene reset mechanism in the enhancement phase to continuously explore potential optimal solutions under the instability of the rescue site environment, thereby improving path efficiency and rescue effectiveness.

[0055] Example 8, see Figure 2 Based on the above embodiments, the Internet of Things-based disaster reduction and relief path planning system provided by the present invention includes an environment interaction module, a network update module, a reward function design module, an inner network training module, an outer optimization search module, and a disaster reduction and relief path planning module.

[0056] The environmental interaction module uses IoT sensors and drones to collaboratively construct a real-time environmental grid map and issue actions in segments to achieve environmental interaction;

[0057] The network update module uses Q network and policy network at different frequencies, combined with a dual-Q structure and replay buffer, to asynchronously update network parameters;

[0058] The reward function design module designs a reward function based on multiple objectives related to rescue needs;

[0059] The inner network training module collects environmental interaction data and continuously iterates the training network parameters to obtain performance indicators.

[0060] The outer-layer optimization search module optimizes hyperparameters, employs adaptive crossover and mutation mechanisms, and uses phased enhancement to filter for the optimal configuration.

[0061] The disaster relief and rescue route planning module deploys the optimal strategy to the disaster area and executes the rescue route planning.

[0062] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0063] 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.

[0064] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A disaster mitigation and relief route planning method based on the Internet of Things, characterized in that: The method includes the following steps: Step S1: Environmental Interaction; Real-time environmental grid map is constructed through collaboration between IoT sensors and drones, and actions are sent out in segments to achieve environmental interaction; Step S2: Network update; Using Q network and policy network at different frequencies, combined with double-Q structure and replay buffer, the network parameters are updated asynchronously; Step S3: Reward function design; Design a reward function based on multiple objectives of rescue needs; Step S4: Inner network training; collect environmental interaction data and continuously iterate the training of network parameters to obtain performance indicators; Step S5: Outer layer optimization search; by optimizing hyperparameters, an adaptive crossover and mutation mechanism is adopted, as well as a phased enhancement to select the optimal configuration; Step S6: Disaster mitigation and relief route planning; Deploy the optimal strategy to the disaster area and execute the relief route planning; In step S3, the reward function design is based on the search and rescue mission, refining the reward function into a multi-objective combination; a temporary priority index is introduced. The reward function is defined piecewise as follows: ; ; ;in, It is a communication link quality indicator; It represents the remaining battery percentage; It is the Euclidean distance between the current point and the target point; It is the distance from the target point in one step; It is the horizontal angle between the drone's current heading and the target direction; It is a communication weighting coefficient; It is the energy weighting coefficient; It is a positive reward for reaching the target area; It is a negative reward for collisions; , and These refer to reaching the target area, collision, and other situations, respectively. It is the distance progress reward coefficient; d is the heading deviation penalty coefficient; d is the relative distance-to-schedule ratio. It represents the remaining target points; It is the maximum number of target points; It is the task threshold; sign(·) is the sign function; In step S4, the inner network training involves initializing the cloud computing nodes and cloud server, loading the Q network and policy network parameters; initializing the replay cache, setting the hyperparameters of learning rate, discount factor and update frequency; preparing the UAV and IoT sensor network, and starting to collect real-time environmental data; repeating the iteration until convergence or reaching the maximum number of rounds, and the network training is completed; the iteration content is: (1) the current policy network is based on the state Give an action , will the action Break it down into steering, forward, and elevation / reclining actions, with delay. The segmentation strategy is distributed to the drone; by frequency Monitor the drone's status and surrounding environment, and record collision warnings in a timely manner; using frequency (1) Record an interaction sample to the replay buffer; (2) If the Q-network or policy network update condition is met, randomly sample a batch of samples from the replay buffer; update the target using the minimum value of the dual Q-network and calculate the target value. And perform gradient descent on the online Q network parameters; use the updated Q network to guide the gradient ascent of the policy network and update the policy network parameters; after the synchronization condition is reached, copy the online Q network parameters to the target network; (3) continue to execute the next time step until the end of the whole round; at this time, the single inner layer training ends, and the performance index under this hyperparameter configuration is output; performance index Represented as: ;in, , , and These are average cumulative reward, collision rate, communication link stability, and energy consumption. , , and These are the corresponding weights; In step S5, the outer layer optimization search specifically includes the following steps: Step S51: The learning rate, discount factor, update frequency, and reward weight are set as individual chromosome parameters, and optimization evaluation is performed in parallel on the cloud. For each generation of optimized individuals, the inner network is trained in parallel on the cloud to obtain the fitness of each individual in the population. M individuals are initialized and mapped to a binary population. Performance metrics are used as individual fitness. Crossover and mutation are performed on individuals based on adaptive probabilities. For the nth generation, the crossover and mutation probabilities of individuals are expressed as: ; ;in, It is the probability that an individual performs a crossover operation; and These are the upper and lower bounds of the crossover probability, respectively. It is the probability that an individual will perform a mutation operation; and These are the upper and lower bounds of the mutation operation, respectively; It is the average fitness value of all individuals in the current generation; It is the maximum fitness value of all individuals in the current generation; It is the maximum fitness among the two parents involved in the crossover; It is the individual's own adaptability; It is the maximum iteration algebra; It is an exponential decay factor; Step S52: Reset enhancements; every m generations, select the middle-ranked fitness groups. Individual genes are randomly reset, and the crossover / mutation probability is temporarily increased, as shown below: ; ;in, and These are the crossover and mutation probabilities after the temporary boost to individuals during the reset enhancement phase, respectively; w is the window center. It is the shape factor; Step S53: Set a threshold and end when the maximum number of generations is reached or the fitness fluctuation is less than the threshold; obtain the hyperparameters of the optimal individual representation; send the hyperparameters of the optimal individual representation to the cloud node, replace the learning rate, discount factor, update frequency and reward weight settings, and train the Q network and policy network.

2. The disaster mitigation and relief route planning method based on the Internet of Things according to claim 1, characterized in that: In step S1, the environmental interaction involves constructing a real-time environmental grid map using IoT sensors and drones deployed in the disaster area; decomposing the action signals output by the path planning strategy network into turning, forward movement, and ascent / descent; firstly, a turning command is issued to the drone from the cloud computing node; then, a delay is applied. Then issue forward and vertical commands; Throughout the entire execution of the action, at frequency Acquire status data from IoT sensor networks and UAVs for collision warning; UAV status includes position, speed, and obstacle distance; at low frequency. Record the environment state, strategy actions, and execution results as a complete interaction experience and store it in the cloud experience playback cache.

3. The disaster mitigation and relief route planning method based on the Internet of Things according to claim 2, characterized in that: In step S2, the network update involves storing accumulated experience in the replay cache of cloud nodes, randomly sampling it in batches, and using it for updating the Q network and policy network. A delayed cyclic step strategy is employed, with the Q network and policy network updating at different frequencies. Two Q networks are deployed in the cloud: an online network and a target network. The online network updates in real time, while the target network updates every [time period missing]. The parameters are synchronized once per iteration; each Q-network update minimizes the deviation by using the smaller of the two outputs.

4. The disaster mitigation and relief route planning method based on the Internet of Things according to claim 3, characterized in that: In step S6, the disaster relief path planning involves deploying a pre-trained Q network and policy network at the disaster site; based on the real-time status, the policy network is directly invoked to output the optimal action; a segmented execution mechanism and IoT sensing information are also used to maintain stable control and collision warning for the drone.

5. An Internet of Things (IoT)-based disaster mitigation and relief route planning system, used to implement the IoT-based disaster mitigation and relief route planning method as described in any one of claims 1-4, characterized in that: It includes an environment interaction module, a network update module, a reward function design module, an inner network training module, an outer optimization search module, and a disaster relief path planning module; The environmental interaction module uses IoT sensors and drones to collaboratively construct a real-time environmental grid map and issue actions in segments to achieve environmental interaction; The network update module uses Q network and policy network at different frequencies, combined with a dual-Q structure and replay buffer, to asynchronously update network parameters; The reward function design module designs a reward function based on multiple objectives related to rescue needs; The inner network training module collects environmental interaction data and continuously iterates the training network parameters to obtain performance indicators. The outer-layer optimization search module optimizes hyperparameters, employs adaptive crossover and mutation mechanisms, and uses phased enhancement to filter for the optimal configuration. The disaster relief and rescue route planning module deploys the optimal strategy to the disaster area and executes the rescue route planning.

Images