Unmanned aerial vehicle cluster resource scheduling method for dynamic task scene

Abstract

The invention discloses a dynamic task scene-oriented unmanned aerial vehicle cluster resource scheduling method, which relates to the technical field of unmanned aerial vehicles, and receives task instruction change in real time through a sensor network and an edge computing node, updates a global dynamic task atlas and transmits the global dynamic task atlas to a central control node for processing. And the central control node decomposes the task instruction to generate a sub-task packet, assesses the global state change of resource allocation, and triggers a rescheduling decision or issues the sub-task packet. And after the distributed nodes receive the sub-task packets, the task emergency degree is judged according to the task deadline, and resource scheduling is performed by adopting an improved genetic algorithm or model prediction control. Meanwhile, the state of the central control node is detected through a heartbeat mechanism, and a distributed negotiation mode is adopted for fault tolerance when failure occurs, so that the problem that limitation exists when a traditional centralized scheduling method and a traditional distributed scheduling method deal with rapidly changing task requirements and real-time requirements is solved; according to the invention, efficient resource scheduling of the unmanned aerial vehicle cluster in a complex dynamic scene is realized.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and more specifically to a method for scheduling UAV swarm resources for dynamic mission scenarios. Background Technology

[0002] With the rapid development of drone technology, drone swarms have demonstrated enormous application potential in various fields such as search and rescue, disaster assessment, environmental monitoring, and agricultural plant protection. Drone swarms, composed of a large number of drones, can perform large-scale, multi-objective tasks through collaborative work, exhibiting strong adaptability. This swarm-based working method has become an important trend in the future development of drones and is frequently applied in dynamic and complex mission scenarios. In dynamic mission scenarios, the tasks of drone swarms are characterized by suddenness and complexity, with mission requirements and priorities changing constantly, posing numerous challenges to their operation. Therefore, resource scheduling is particularly important to enable drone swarms to execute tasks efficiently in real-world environments.

[0003] In existing technologies, commonly used UAV swarm resource scheduling methods are centralized scheduling and distributed scheduling. Centralized scheduling uses a central controller to uniformly plan task allocation and resource scheduling. It requires collecting and processing the status information and task requirements of all UAVs. In dynamic task scenarios, task requirements and priorities change at any time. The central controller has high computational and communication overhead, resulting in slow response speed, poor real-time performance, and difficulty in adapting to rapidly changing task requirements. Distributed scheduling relies on local information exchange for autonomous decision-making. UAVs need to communicate and negotiate multiple times to reach a consensus. In dynamic task scenarios, task requirements and priorities change frequently, and the algorithm convergence time is long, which cannot meet real-time requirements.

[0004] Therefore, a method for scheduling drone swarm resources for dynamic task scenarios is needed to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention discloses a resource scheduling method for UAV swarms in dynamic task scenarios, which can achieve efficient and flexible task scheduling and resource allocation, thereby improving the task execution efficiency and reliability of UAV swarms.

[0006] The present invention adopts the following technical solution:

[0007] A method for scheduling drone swarm resources for dynamic task scenarios includes the following steps:

[0008] S1. Receive dynamic changes in task instructions in real time through sensor networks and edge computing nodes, update the global dynamic task map and transmit it to the central control node for preliminary processing.

[0009] S2. The central control node decomposes the task instruction into sub-task packages based on the sudden time, target location, resource requirements and deadline window. It also evaluates the global state change value of the UAV cluster resource allocation after adding the sub-task package based on the resource requirement gap rate, task urgency increment and execution time-space overlap rate. When the global state change value exceeds the preset threshold, the UAV cluster resource rescheduling decision is triggered. Otherwise, the sub-task package is sent to the distributed nodes of the UAV cluster for local resource scheduling and task execution.

[0010] S3. After receiving the sub-task package issued by the central control node, the distributed nodes of the UAV cluster determine that the sub-task package is an emergency task if the task deadline is less than a preset threshold, otherwise determine that the sub-task package is a regular task. When the sub-task package is determined to be an emergency task, an improved genetic algorithm is used to make the optimal resource scheduling decision. When the sub-task package is determined to be a regular task, model predictive control is used to make dynamic resource scheduling.

[0011] S4. During the resource scheduling process of the drone cluster, the status of the central control node is detected in real time through a heartbeat mechanism. When the failure of the central control node is detected, a distributed negotiation mode is adopted for fault tolerance.

[0012] Furthermore, in S1, the edge computing node uses Kalman filtering to eliminate the spatiotemporal drift of sensor network data, and extracts information on burst time, target location, resource requirements and deadline window of received task instructions based on a lightweight self-supervised learning engine. Then, a global dynamic task graph is constructed based on the extracted information. The global dynamic task graph uses a spatiotemporal grid as primitives and is updated using spatiotemporal differential coding.

[0013] Furthermore, in S2, when the rescheduling decision of the UAV cluster resources is triggered, the central control node dynamically calculates the execution priority of the subtask package using a multi-factor fusion priority model, makes a global resource scheduling decision based on the execution priority of the subtask package, and distributes the subtask package to the distributed nodes of the UAV cluster for local resource scheduling and execution based on the generated global resource scheduling strategy.

[0014] Furthermore, the multi-factor fusion priority model calculates the execution priority of sub-task packages by linearly weighting the urgency of the task, the resource matching degree, and the environmental risk factor, and adopts a preemption mechanism to execute sub-task packages in descending order of priority. When making global resource scheduling decisions, the central control node abstracts the capabilities of the UAV swarm into resource units to be allocated, and uses a consistent hashing algorithm to match sub-task packages with resource units to be allocated. The capabilities of the UAV swarm include at least flight area, flight time, payload capacity, communication capability, and sensor capability.

[0015] Furthermore, in S3, after receiving the sub-task package from the central control node, the distributed nodes of the UAV cluster dynamically adjust the execution strategy of the sub-task package according to the UAV status, mission progress and environmental obstacles. When the distributed nodes of the UAV cluster are abnormal, a compensation mechanism is used to transfer the sub-task package to a nearby idle distributed node. The compensation mechanism dynamically reorganizes the UAV formation according to mission requirements and realizes the hot migration of the sub-task package between different distributed nodes of the UAV cluster through containerization.

[0016] Furthermore, in S3, when a subtask package is determined to be an urgent task, the improved genetic algorithm operates by including the following steps:

[0017] S301. Receive the list of urgent subtask packages and randomly generate a set of scheduling schemes as the initial solution, and dynamically update the encoding method according to the urgency of the task.

[0018] S302. Each scheduling scheme in the initial solution is evaluated for fitness based on the dimensions of emergency task completion rate, resource utilization rate and conflict penalty, and the priority of emergency tasks is maintained by using the task urgency weight parameter.

[0019] S303. Perform selection, crossover, and mutation operations on each scheduling scheme in the initial solution;

[0020] S304. Repeat S302 to S303 to iteratively optimize the scheduling scheme. In each iteration, select the scheduling scheme as the parent scheme based on the fitness evaluation result and generate a new child scheduling scheme. When the maximum number of iterations is reached, output all non-dominated solutions.

[0021] S305. The rule engine is used to select the optimal solution from the non-dominated solutions and distribute it to the distributed nodes of the drone cluster for local resource scheduling and task execution.

[0022] Furthermore, in S3, when a subtask package is determined to be a regular task, a lightweight digital twin model is used to predict the task execution environment, task requirements, and resource status. Based on the prediction results, the resource scheduling scheme is continuously optimized. Then, the resource scheduling scheme is fed back and corrected by real-time monitoring of task execution and resource status.

[0023] Furthermore, in S4, the heartbeat mechanism uses an adaptive heartbeat protocol to send detection signals from the distributed nodes of the UAV cluster to the central control node. After receiving the detection signal, the central control node replies with an acknowledgment signal to prove that the central control node is in normal condition; otherwise, the central control node is in failure. In the distributed negotiation mode, the distributed nodes of the UAV cluster communicate and negotiate with each other through a lightweight communication protocol.

[0024] The beneficial effects of this invention are as follows:

[0025] 1. This invention adopts a layered architecture, including a sensor network, edge computing nodes, a central control node, and distributed nodes in a drone swarm. This makes the responsibilities of each component clear, facilitating independent development, testing, and maintenance. The sensor network and edge computing nodes are responsible for data acquisition and preliminary processing, the central control node is responsible for global task decomposition and resource status assessment, and the distributed nodes in the drone swarm are responsible for local resource scheduling and task execution. This layered architecture improves the system's scalability and flexibility.

[0026] 2. This invention employs a central control node for global task decomposition and resource status assessment. It dynamically calculates the execution priority of sub-task packages using a multi-factor fusion priority model, comprehensively considering multiple factors affecting task execution. This ensures tasks are rationally scheduled and executed according to their importance and urgency. When the global status change value exceeds a preset threshold, the central control node triggers a rescheduling decision for the UAV swarm resources. This dynamic scheduling decision mechanism can adjust resource allocation promptly based on changes in tasks and the environment, improving task adaptability and robustness. When the global status change value is below the preset threshold, it is directly distributed to the distributed nodes of the UAV swarm for local resource scheduling and task execution, improving task execution efficiency.

[0027] 3. This invention employs distributed nodes for local resource scheduling and task execution. Based on a comparison of task deadlines with preset thresholds, subtask packages are dynamically categorized into urgent and routine tasks. This categorization mechanism allows distributed nodes to adopt different scheduling strategies for different task types, improving the flexibility and efficiency of task execution. For urgent tasks, distributed nodes use an improved genetic algorithm for optimal resource scheduling decisions. Upon receiving a subtask package, distributed nodes can dynamically adjust the subtask package execution strategy based on the UAV's status, task progress, and environmental obstacles. This dynamic adjustment mechanism enables the UAV swarm to better adapt to complex and changing task environments, improving the reliability and success rate of task execution. When a distributed node malfunctions, a compensation mechanism is used to transfer subtask packages to nearby idle distributed nodes. This compensation mechanism ensures that tasks can continue to execute even when a node fails, improving the system's robustness and reliability.

[0028] 4. This invention employs a central control node for global task planning and macro-level resource scheduling, while using distributed nodes for local scheduling and execution of specific tasks. This combined global and local scheduling approach ensures both global optimization and local flexibility and efficiency, avoiding bottlenecks caused by relying solely on centralized or distributed resource scheduling. This results in more efficient and stable task processing. When the central node fails, the system switches to a distributed negotiation mode to ensure the continuous operation of the drone swarm, thereby improving the efficiency and stability of task execution. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0030] Figure 2 This is a flowchart illustrating the improved genetic algorithm in this invention. Detailed Implementation

[0031] The following will refer to the appendices in the embodiments of the present invention. Figure 1 To be continued Figure 2 The technical solutions in the embodiments of the present invention are clearly and completely described herein. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0032] This invention discloses a method for scheduling unmanned aerial vehicle (UAV) swarm resources for dynamic task scenarios, as shown in the attached figure. Figure 1 This includes the following steps:

[0033] S1. Construction and updating of the global dynamic task graph

[0034] The system receives dynamic changes in task instructions in real time through sensor networks and edge computing nodes, updates the global dynamic task map, and transmits it to the central control node for preliminary processing.

[0035] Deploy a sensor network containing visual sensors, infrared sensors, cameras, and GPS multimodal sensors to capture dynamic changes in the task in real time, including information on emergencies, resource requirements, and environmental disturbances, and transmit this information to edge computing nodes via the MQTT / CoAP protocol.

[0036] Edge computing nodes use Kalman filtering to eliminate spatiotemporal drift, solving the problem of spatiotemporal drift caused by sensor clock asynchrony and positioning errors. The state vector of the filter includes position, velocity and timestamp deviation, and is optimized through prediction-correction loop iteration.

[0037] Edge computing nodes extract burst time, target location, resource requirements, and deadline window information of task instructions based on a lightweight self-supervised learning engine. This lightweight self-supervised learning engine, based on a Temporal CNN encoder-decoder architecture, reconstructs the time-series data of pre-trained task instructions through masking, learns the implicit patterns of features such as burst time and resource requirements, and outputs structured task burst time, target location, resource requirements, and deadline window as node attributes in the task graph. The information extraction process includes:

[0038] Emergency Time Extraction: Analyze time-descriptive words in the task instruction text and combine them with time context information to determine the emergency time of the task. For example, by recognizing words such as "immediately" and "at [specific time]", the start time of the task can be extracted;

[0039] Target location extraction: Utilizing natural language processing techniques, such as named entity recognition, to identify entity information of the target location from task instructions. For example, identifying location names such as "airport" and "warehouse," and combining this with Geographic Information System (GIS) data to determine the target's specific location coordinates;

[0040] Resource requirement extraction: Analyze the resource descriptions in the task instructions, such as the number of drones and payload type. Resource requirement information can be extracted using keyword matching and semantic analysis.

[0041] Deadline window extraction: Determine the time range within which the task needs to be completed. Extract deadline window information by analyzing time constraint words in the task instructions, such as "complete before [time 1]" and "complete within [time 2]-[time 3]".

[0042] Edge computing nodes construct a spatiotemporally gridded global dynamic task map based on extracted information. This global dynamic task map uses a spatiotemporal grid as its primitive, discretizing the task space into a three-dimensional spatiotemporal grid. This grid includes longitude, latitude, and time slots. Each grid stores statistics such as task density and total resource requirements. The spatiotemporal grid primitives divide the time axis into several discrete time slices based on the task's dynamic characteristics and time granularity requirements. For example, for tasks with high real-time requirements, a time slice is set to 1 second; for tasks with lower real-time requirements, a time slice is set to 1 minute. The spatial region of the task scenario is also divided into regular grids. The grid size is adjusted according to the task's accuracy requirements and computing resources. For example, in urban environmental monitoring tasks, the urban area is divided into 100m × 100m grids. Differential encoding is used to transmit only grid data that changes between adjacent time points, reducing communication overhead. The update cycle is dynamically adjusted based on the urgency of the deadline window, integrating scattered task information into a global view, providing a foundation for subsequent decomposition and scheduling. Spatiotemporal differential coding uses binary coding or other efficient coding methods to record the state changes of a spatiotemporal grid. For example, 0 represents that the grid is idle at a certain moment, and 1 represents that the grid is occupied by a task. When a new task appears, only the coding of the affected spatiotemporal grid is updated. For example, when the target location of a new task is located within a certain grid and the task's time window covers a certain time slice, the coding of that grid in the corresponding time slice is updated.

[0043] Data transmission protocols, such as TCP / IP, are used to ensure the accurate and error-free transmission of global dynamic task map data to the central control node. During transmission, the data is compressed to reduce transmission bandwidth and time.

[0044] S2, The central control node makes decisions on task decomposition and resource reallocation.

[0045] Once the central control node receives a task instruction, it analyzes the instructions in the global dynamic task map, extracting key information such as the time of the incident, target location, resource requirements, and deadline window. Based on these factors, the task instruction is broken down into multiple independent sub-task packages. Each sub-task package contains detailed information such as a clear task description, execution time, target location, and required resources. For example, a sub-task package might be generated where the task is to perform a specific type of monitoring at a target location at a specific time, and the required resource is a drone equipped with specific sensors.

[0046] Then, the global state change in drone swarm resource allocation after the addition of a new subtask package is assessed based on the resource demand gap rate, task urgency increment, and execution spatiotemporal overlap rate. The resource demand gap rate is the ratio of the gap between the drone swarm's existing resources (such as the number of drones, payload capacity, etc.) and the resources required for the task. The formula is: Resource demand gap rate = (Task-required resources - Existing swarm resources) / Task-required resources × 100%. For example, if a task requires 10 drones, and the swarm currently has 8 available drones, the resource demand gap rate is 20%. The task urgency increment is used to assess the impact of the new subtask package on the overall task urgency. The change in task urgency is calculated based on the task's deadline window and contingency time. For example, if the deadline for the new task is very tight, it will significantly increase the overall task urgency. The urgency increment can be calculated based on the proximity of the deadline to the current time by setting an urgency scoring standard. The execution spatiotemporal overlap rate is used to analyze the temporal and spatial overlap between the new subtask package and existing tasks. By calculating the temporal and spatial overlap ratio, the degree of conflict between tasks is assessed. For example, if two tasks are executed at the same time and in the same region, the spatiotemporal overlap rate is 100%. The overlap rate can be calculated by comparing the time intervals and spatial ranges of the tasks.

[0047] If the global state change value after comprehensive evaluation exceeds a preset threshold, the system will trigger resource rescheduling. Otherwise, the subtask package will be directly issued to the distributed nodes of the drone cluster for local scheduling. The global state change value is obtained by weighted summation of the resource demand gap rate, task urgency increment, and spatiotemporal overlap rate. The formula is: Global State Change Value = Resource Demand Gap Rate × Weight 1 + Task Urgency Increment × Weight 2 + Execution Spatiotemporal Overlap Rate × Weight 3. The weights can be adjusted according to the actual situation. For example, for urgent task scenarios, the weight of the task urgency increment can be set higher. When the global state change value exceeds this threshold, it indicates that the current resource allocation needs to be significantly adjusted, triggering a rescheduling decision for drone cluster resources; otherwise, the subtask package is issued to the distributed nodes of the drone cluster for local resource scheduling and task execution.

[0048] When a rescheduling decision for drone cluster resources is triggered, the central control node uses a multi-factor fusion priority model to dynamically calculate the execution priority of the sub-task package, makes a global resource scheduling decision based on the execution priority of the sub-task package, and distributes the sub-task package to the distributed nodes of the drone cluster for local resource scheduling and execution based on the generated global resource scheduling policy.

[0049] The multi-factor fusion prioritization model calculates the execution priority of sub-task packages by linearly weighting the urgency, resource matching degree, and environmental risk factors of the task. A preemptive mechanism is then used to execute sub-task packages in descending order of priority. The model determines the weights of urgency, resource matching degree, and environmental risk factors based on the importance of the task and the actual situation of the drone swarm. For example, the weight of the environmental risk factor is increased when executing a task in a hazardous environment. These weights are determined through historical data analysis. For each sub-task package, its urgency, resource matching degree, and environmental risk factor scores are calculated separately, and then linearly weighted and summed according to their weights to obtain the execution priority of the sub-task package. The formula is: Execution Priority = Urgency Score × Weight 1 + Resource Matching Degree Score × Weight 2 + Environmental Risk Factor Score × Weight 3. Urgency is scored based on factors such as the task's deadline window and the imminent nature of emergencies; tasks with tighter deadlines and closer emergencies receive higher scores. The resource matching degree score is obtained by assessing the degree of matching between the resources required by the sub-task package and the existing resources of the drone swarm; a higher matching degree results in a higher score. For example, if the payload capacity of a subtask package perfectly matches the payload capacity of a drone in the cluster, then the resource matching score for that subtask package on that drone is high. The environmental risk factor score considers environmental factors in the task execution area, such as weather conditions and terrain complexity; the higher the environmental risk, the lower the score.

[0050] Based on the calculated execution priority, a preemptive mechanism is used to execute subtask packages in descending order of priority. When a high-priority subtask package appears, if resources are already occupied by a low-priority subtask package, the execution of the low-priority subtask package can be paused or adjusted to prioritize the execution of the high-priority subtask package. For example, if a high-priority emergency rescue mission and a low-priority monitoring mission both require the same drone, the monitoring mission is paused, and the drone performs the rescue mission.

[0051] When making global resource scheduling decisions, the central control node abstracts the capabilities of the UAV swarm into resource units to be allocated. For example, a UAV's flight area, flight time range, maximum payload, communication bandwidth, and sensor type are considered as a resource unit. A consistent hashing algorithm is used to match subtask packages with resource units to be allocated. The capabilities of the UAV swarm include at least flight area, flight time, payload capacity, communication capacity, and sensor capacity. UAV nodes and subtask packages are mapped to a virtual ring with 2^32 buckets using a hash function. For each subtask package, its hash value is calculated and mapped to the hash ring, then the nearest UAV node is searched clockwise for task allocation. This ensures even distribution of tasks within the UAV swarm, preventing some UAVs from being overloaded while others are idle. Through hash ring mapping and clockwise search, tasks and UAV resources are quickly matched, improving scheduling efficiency. Dynamic addition and removal of UAV nodes are supported to adapt to changes in swarm size. In the event of node failure or network fluctuations, task allocation remains relatively stable, improving system reliability.

[0052] Based on the generated global resource scheduling strategy, subtask packages are distributed to the distributed nodes of the UAV cluster. During the distribution process, the accuracy of the subtask package information is ensured, including the task objective, execution time, and required resources. After receiving the subtask package, the distributed nodes perform local resource scheduling and task execution.

[0053] Example 1

[0054] The central control node receives an instruction containing multiple monitoring and rescue tasks. Based on information such as the time of the emergency and the location of the target, these tasks are broken down into multiple sub-task packages. For example, sub-task package A is to conduct environmental monitoring of [Area 1] at [Time 1], and sub-task package B is to conduct a rescue mission in [Area 2] at [Time 2].

[0055] When calculating the resource demand gap rate, it was found that the current drone swarm's number of drones is insufficient to meet the needs of all sub-task packages, with a gap rate of 30%. Assessing the mission urgency increment, the overall mission urgency increased by 20% due to the extremely tight deadline for the rescue mission. Calculating the execution spatiotemporal overlap rate revealed that sub-task packages A and B have some overlap in time and space, with an overlap rate of 10%.

[0056] If the global state change value is calculated comprehensively, and the preset threshold is 50%, while the actual calculated global state change value is 60%, exceeding the threshold, a rescheduling decision is triggered.

[0057] The execution priority of sub-task packages was calculated using a multi-factor fusion priority model. Sub-task package B, being a rescue mission, had a higher urgency score, better resource matching, and relatively lower environmental risk, resulting in an execution priority of 0.85. Sub-task package A had an execution priority of 0.65.

[0058] According to the preemption mechanism, subtask package B is executed first. When making global resource scheduling decisions, the resources of the UAV cluster are abstracted into resource units. A consistent hashing algorithm is used to match subtask package A and subtask package B to appropriate UAV resource units, and then the subtask packages are distributed to the distributed nodes of the UAV cluster for execution.

[0059] S3, Distributed nodes make local resource scheduling decisions.

[0060] After receiving a subtask packet from the central control node, the distributed nodes of the drone swarm retrieve the task deadline information from the subtask packet. They then compare the deadline with a preset threshold to determine whether the subtask packet is an urgent or regular task. The preset threshold is a time value pre-set based on the importance of the task and the system's tolerance range for urgency. If the task deadline is less than the preset threshold, the subtask packet is determined to be an urgent task; otherwise, it is determined to be a regular task. For example, if the preset threshold is 1 hour, and the deadline of the subtask packet is 30 minutes later, then the subtask packet is determined to be an urgent task.

[0061] For urgent tasks, an improved genetic algorithm is used to make optimal resource scheduling decisions. This improved genetic algorithm optimizes the traditional genetic algorithm, enabling it to converge to the optimal solution more quickly and improving the efficiency and accuracy of resource scheduling. The working principle of the improved genetic algorithm is shown in the attached diagram. Figure 2 As shown, it includes the following steps:

[0062] S301: Initial Solution Generation and Encoding Update

[0063] The system receives a list of urgent subtask packages and randomly generates a set of scheduling schemes as the initial solution. Each scheduling scheme represents a way of allocating drones and tasks. The encoding method is dynamically updated according to the urgency of the tasks. For example, for tasks with higher urgency, a more refined encoding method can be used to more accurately represent the task allocation.

[0064] S302: Fitness Assessment and Priority Maintenance

[0065] Each scheduling scheme in the initial solution is evaluated for fitness based on three dimensions: emergency task completion rate, resource utilization rate, and conflict penalty. The emergency task completion rate reflects the scheduling scheme's ability to complete emergency tasks; resource utilization rate measures the effective use of resources; and conflict penalty considers resource conflicts between UAVs. An emergency task weighting parameter is used to maintain the priority of emergency tasks. In the fitness evaluation, higher weights are assigned to indicators related to emergency tasks to ensure that emergency tasks are given priority in the scheduling schemes.

[0066] S303: Selection, crossover, and mutation operations

[0067] Based on fitness assessment results, a superior scheduling scheme is selected as the parent generation to generate new offspring scheduling schemes. Common selection methods include roulette wheel selection and tournament selection. A crossover operation is performed on the parent scheduling scheme to generate new offspring scheduling schemes. The crossover operation can exchange some genes in the parent scheme, producing new combinations. A mutation operation is then performed on the offspring scheduling scheme to increase population diversity. The mutation operation can randomly change certain gene values ​​in the scheduling scheme.

[0068] S304: Iterative Optimization and Non-dominated Solution Output

[0069] Repeat steps S302 to S303 to iteratively optimize the scheduling scheme. In each iteration, select a scheduling scheme as the parent scheme based on the fitness evaluation results and generate a new child scheduling scheme. When the maximum number of iterations is reached, output all non-dominated solutions. A non-dominated solution is one in which no single scheduling scheme is superior to all other schemes in all dimensions.

[0070] S305: Optimal Solution Selection and Distribution

[0071] A rule engine is used to select the optimal solution from non-dominated solutions. The rule engine can select the most suitable scheduling scheme based on preset rules (such as the highest completion rate of emergency tasks, the highest resource utilization, etc.).

[0072] For routine tasks, Model Predictive Control (MMCC) is employed for dynamic resource scheduling. MMCC predicts future resource demands based on dynamic changes in the task and adjusts accordingly to achieve rational resource allocation and utilization. It uses a lightweight digital twin model to predict the task execution environment, task requirements, and resource status. This prediction forecasts changes in the task execution environment, dynamic adjustments to task requirements, and changes in resource status over a future period. The digital twin model is a virtual mapping of the physical system, simulating the system's operational state in real time. The resource scheduling scheme is then continuously optimized based on the prediction results. At each sampling time, an optimization problem within a finite time domain is solved based on the current prediction results to obtain the optimal resource scheduling scheme. The resource scheduling scheme is corrected by real-time monitoring of task execution and resource status. When deviations occur between the actual execution and the prediction results, the resource scheduling scheme is adjusted promptly to ensure successful task execution.

[0073] After receiving sub-task packages from the central control node, the distributed nodes of the UAV swarm dynamically adjust the execution strategy of the sub-task packages based on the UAV status, mission progress, and environmental obstacles. When a distributed node in the UAV swarm malfunctions, a compensation mechanism is used to transfer the sub-task packages to a nearby idle distributed node. This compensation mechanism dynamically reorganizes the UAV formation according to mission requirements and achieves hot migration of sub-task packages between different distributed nodes in the UAV swarm through containerization.

[0074] Upon receiving a subtask package, the distributed nodes of the drone swarm dynamically adjust the execution strategy based on the drone's status, mission progress, and environmental obstacles. When a drone malfunctions or encounters adverse conditions, the execution path or method is adjusted promptly to ensure successful mission completion. If a drone's battery is low, the distributed nodes may reassign the subtask package it was originally executing to another drone with sufficient battery power. If the mission is behind schedule, resources allocated to the mission may be increased, such as adding more drones or increasing their flight speed. When encountering environmental obstacles, the drone's flight path may be adjusted to avoid obstacles, and the execution order and timing of the subtask packages may be replanned.

[0075] The distributed nodes of the drone swarm monitor their own operational status in real time. When an anomaly is detected in a distributed node, a compensation mechanism is used to transfer subtask packages to a nearby idle distributed node. This compensation mechanism dynamically reorganizes the drone formation based on task requirements. If a drone under the responsibility of a distributed node fails, a suitable drone is selected from nearby idle drones to re-form the formation, ensuring the task can continue. Hot migration of subtask packages between different distributed nodes in the drone swarm is achieved through containerization. Containerization technology packages the subtask package and its dependencies into an independent container, and packages the relevant code, data, and configuration files into a container image. When an anomaly is detected in a distributed node, the container image is migrated from the faulty node to a nearby idle node. The container is then started on the idle node, resuming the execution of the subtask package. This facilitates rapid migration and deployment between different nodes, improving the system's fault tolerance and flexibility.

[0076] Example 2

[0077] Suppose that the distributed nodes of a drone swarm receive multiple sub-task packages from the central control node. Sub-task package C has a deadline of 40 minutes, with a preset threshold of 1 hour. Sub-task package C is then classified as an urgent task, and an improved genetic algorithm is used for resource scheduling. Sub-task package D has a deadline of 2 hours, and is classified as a regular task, using model predictive control for dynamic resource scheduling.

[0078] During execution, it was discovered that a drone executing subtask package C had insufficient power. The distributed nodes dynamically adjusted the execution strategy based on the drone's status, reassigning part of the task originally being executed by that drone to other drones with sufficient power.

[0079] If a distributed node fails, the compensation mechanism will select drones from nearby idle drones to re-form the formation according to the task requirements, and migrate the sub-task package from the failed node to the idle node in a containerized manner to ensure that the sub-task package can continue to execute.

[0080] S4, Central Control Node Fault Tolerance

[0081] During the resource scheduling process of a drone swarm, the distributed nodes of the drone swarm periodically send detection signals to the central control node according to an adaptive heartbeat protocol. The adaptive heartbeat protocol dynamically adjusts the sending frequency of heartbeat signals based on real-time network conditions (network latency, packet loss rate, etc.) to reduce network overhead. When network conditions are good, the sending frequency is reduced to decrease network load; when network conditions are poor, the sending frequency is increased to improve detection timeliness. Upon receiving the detection signal, the central control node replies with an acknowledgment signal to the distributed nodes. This acknowledgment signal indicates that the central control node is in normal working order. If a distributed node does not receive an acknowledgment signal from the central control node within a specified time, the central control node is considered to have failed.

[0082] When a central control node failure is detected, a distributed negotiation mode is employed for fault tolerance to ensure the normal operation of the drone swarm. In this mode, the distributed nodes of the drone swarm communicate and negotiate using a lightweight communication protocol. This protocol features low bandwidth consumption and low latency, enabling efficient information transmission even in resource-constrained environments. The distributed nodes negotiate how to reallocate tasks originally managed by the central control node. Each node proposes its own task allocation plan based on its resource status (remaining battery power, computing power, etc.) and task requirements. To ensure the orderly conduct of the distributed negotiation, a leader node needs to be elected. The leader node is responsible for coordinating communication and task allocation among the nodes. Various strategies are used to elect the leader node, such as comprehensively evaluating nodes based on remaining battery power, computing power, and communication quality, selecting the optimal node as the leader. Under the coordination of the leader node, each node adjusts its resource scheduling strategy according to the new task allocation plan. For example, drones originally responsible for a certain area may adjust their flight routes and task execution order based on new task requirements.

[0083] The asynchronous backtracking search algorithm (ABT) is used to resolve task-resource allocation conflicts. Pareto optimality is achieved by maximizing the utility function. The future trajectory of the UAV is predicted based on the kinematic model. The spatiotemporal cube collision detection algorithm (STC) is used to plan avoidance paths in advance. Macro-level task allocation and conflict resolution are performed, task priorities are calculated, and a deep reinforcement learning (DRL) module is embedded. Resource matching strategies and conflict resolution parameters are optimized online using historical scheduling data.

[0084] This approach employs a layered architecture, including a sensor network, edge computing nodes, a central control node, and distributed nodes in a drone swarm. This layered architecture clearly defines the responsibilities of each system component, facilitating independent development, testing, and maintenance.

[0085] Using Gazebo's environment modeling tool, define buildings, roads, obstacles, etc., in the city via SDF files. Construct a realistic urban logistics scenario, including distribution centers, customer locations, traffic lights, etc. Download suitable drone models from the PX4 model library and place them in Gazebo's model directory. Configure necessary sensors for the drones, such as LiDAR, cameras, IMUs, etc., to simulate realistic perception capabilities. Ensure seamless integration between Gazebo and ROS, achieving real-time communication and control through gazebo_ros_pkgs. Utilize ROS's topics, services, and parameter servers to achieve communication and task allocation between drones. During the simulation, randomly generate a large number of delivery orders and dynamically allocate these orders to the drone swarm. Simulate a sudden increase in the number of orders to test the system's performance under high load conditions. Randomly select some drones and simulate their failures (such as battery depletion, sensor malfunction, etc.). Observe how the system reallocates tasks and adjusts resource scheduling strategies to cope with drone failures. An emergency delivery scenario for urban logistics was constructed in Gazebo / ROS, simulating dynamic events such as a surge in orders and drone malfunctions. Comparative experiments showed that this method improved the task completion rate by 21% and reduced energy consumption by 33%. Specific data are shown in Table 1.

[0086]

[0087] The data above shows that this method, through the coordinated work of four steps, can achieve efficient scheduling of UAV swarm resources for dynamic task scenarios, thereby improving the task execution capability and reliability of UAV swarms.

[0088] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, any equivalent modifications or substitutions made by those skilled in the art to the relevant technical features will fall within the scope of protection of the present invention.

Claims

1. A method for scheduling unmanned aerial vehicle (UAV) swarm resources for dynamic task scenarios, characterized in that, Includes the following steps: S1. Receive dynamic changes in task instructions in real time through sensor networks and edge computing nodes, update the global dynamic task map and transmit it to the central control node for preliminary processing. S2. The central control node decomposes the task instruction into sub-task packages based on the sudden time, target location, resource requirements and deadline window. It also evaluates the global state change value of the UAV cluster resource allocation after adding the sub-task package based on the resource requirement gap rate, task urgency increment and execution time-space overlap rate. When the global state change value exceeds the preset threshold, the UAV cluster resource rescheduling decision is triggered. Otherwise, the sub-task package is sent to the distributed nodes of the UAV cluster for local resource scheduling and task execution. S3. After receiving the sub-task package issued by the central control node, the distributed nodes of the UAV cluster determine that the sub-task package is an emergency task if the task deadline is less than a preset threshold, otherwise determine that the sub-task package is a regular task. When the sub-task package is determined to be an emergency task, an improved genetic algorithm is used to make the optimal resource scheduling decision. When the sub-task package is determined to be a regular task, model predictive control is used to make dynamic resource scheduling. S4. During the resource scheduling process of the drone cluster, the status of the central control node is detected in real time through a heartbeat mechanism. When the failure of the central control node is detected, a distributed negotiation mode is adopted for fault tolerance.

2. The method for scheduling UAV swarm resources for dynamic task scenarios according to claim 1, characterized in that, In S1, the edge computing node uses Kalman filtering to eliminate the spatiotemporal drift of sensor network data, and extracts information on burst time, target location, resource requirements and deadline window of received task instructions based on a lightweight self-supervised learning engine. Then, a global dynamic task graph is constructed based on the extracted information. The global dynamic task graph uses a spatiotemporal grid as primitives and is updated using spatiotemporal differential coding.

3. The method for scheduling UAV swarm resources for dynamic task scenarios according to claim 1, characterized in that, In S2, when the rescheduling decision of the UAV cluster resources is triggered, the central control node dynamically calculates the execution priority of the sub-task package using a multi-factor fusion priority model, makes a global resource scheduling decision based on the execution priority of the sub-task package, and distributes the sub-task package to the distributed nodes of the UAV cluster for local resource scheduling and execution based on the generated global resource scheduling strategy.

4. The method for scheduling UAV swarm resources for dynamic task scenarios according to claim 3, characterized in that, The multi-factor fusion priority model calculates the execution priority of sub-task packages by linearly weighting the urgency of the task, resource matching degree, and environmental risk factors, and uses a preemption mechanism to execute sub-task packages in descending order of priority. When making global resource scheduling decisions, the central control node abstracts the capabilities of the UAV swarm into resource units to be allocated, and uses a consistent hashing algorithm to match sub-task packages with resource units to be allocated. The capabilities of the UAV swarm include at least flight area, flight time, payload capacity, communication capability, and sensor capability.

5. The method for scheduling UAV swarm resources for dynamic task scenarios according to claim 1, characterized in that, In S3, after receiving the sub-task package from the central control node, the distributed nodes of the UAV cluster dynamically adjust the execution strategy of the sub-task package according to the UAV status, mission progress and environmental obstacles. When the distributed nodes of the UAV cluster are abnormal, a compensation mechanism is used to transfer the sub-task package to a nearby idle distributed node. The compensation mechanism dynamically reorganizes the UAV formation according to mission requirements and realizes hot migration of sub-task packages between different distributed nodes of the UAV cluster through containerization.

6. The method for scheduling UAV swarm resources for dynamic task scenarios according to claim 1, characterized in that, In S3, when a subtask package is determined to be an urgent task, the improved genetic algorithm operates by including the following steps: S301. Receive the list of urgent subtask packages and randomly generate a set of scheduling schemes as the initial solution, and dynamically update the encoding method according to the urgency of the task. S302. Each scheduling scheme in the initial solution is evaluated for fitness based on the dimensions of emergency task completion rate, resource utilization rate and conflict penalty, and the priority of emergency tasks is maintained by using the task urgency weight parameter. S303. Perform selection, crossover, and mutation operations on each scheduling scheme in the initial solution; S304. Repeat S302 to S303 to iteratively optimize the scheduling scheme. In each iteration, select the scheduling scheme as the parent scheme based on the fitness evaluation result and generate a new child scheduling scheme. When the maximum number of iterations is reached, output all non-dominated solutions. S305. The rule engine is used to select the optimal solution from the non-dominated solutions and distribute it to the distributed nodes of the drone cluster for local resource scheduling and task execution.

7. A method for scheduling unmanned aerial vehicle (UAV) swarm resources for dynamic task scenarios according to claim 1, characterized in that, In S3, when a subtask package is determined to be a regular task, a lightweight digital twin model is used to predict the task execution environment, task requirements, and resource status. Based on the prediction results, the resource scheduling scheme is continuously optimized. Then, the resource scheduling scheme is corrected by real-time monitoring of task execution and resource status.

8. A method for scheduling UAV swarm resources for dynamic task scenarios according to claim 1, characterized in that, In S4, the heartbeat mechanism uses an adaptive heartbeat protocol to send detection signals from the distributed nodes of the UAV cluster to the central control node. After receiving the detection signal, the central control node replies with an acknowledgment signal to prove that the central control node is in normal condition; otherwise, the central control node is in failure. In the distributed negotiation mode, the distributed nodes of the UAV cluster communicate and negotiate with each other through a lightweight communication protocol.

Images