Optimal path selection and defense cooperation method for unmanned aerial vehicle cluster edge calculation

Abstract

The invention discloses an optimal path selection and defense cooperation method for unmanned aerial vehicle cluster edge calculation, and the method comprises the steps: constructing an average field game model based on a composite average field, and designing a function which couples the cost of paths, survival, congestion and calculation. And an iterative learning method is adopted to carry out offline solving: alternately training a strategy network used for generating an individual optimal action and a U-Net structure prediction network used for predicting dynamic evolution of a macroscopic average field until convergence to average field equilibrium. Carrying out distributed online deployment on the strategy; periodically reporting a state to an edge computing node by the unmanned aerial vehicle; the nodes aggregate information and generate a local average field of a future time domain by using a prediction network; the prediction field is issued to the unmanned aerial vehicle, and the unmanned aerial vehicle combines the state of the unmanned aerial vehicle and the prediction field to make a collaborative decision through a local strategy network. According to the invention, integration of large-scale cluster paths, defense and calculation tasks, expandability and predictable intelligent collaboration are realized.

Description

Technical Field

[0001] This invention relates to the field of collaborative control technology for unmanned aerial vehicle (UAV) swarms, and in particular to an optimal path selection and defense collaboration method for edge computing of UAV swarms. Background Technology

[0002] With the rapid development of drone technology, large-scale drone swarms have shown great application potential in fields such as military reconnaissance, collaborative attack and defense, emergency communication, and intelligent logistics due to their significant advantages in distributed perception, collaborative operation, and robustness. To achieve intelligence and autonomy in swarms, multi-agent reinforcement learning (MARL) technology has been widely introduced. By endowing individuals with learning and interaction capabilities, it enables the swarm to exhibit complex collaborative behaviors. Simultaneously, to address the contradiction between limited onboard computing resources and the demand for massive data processing in drones, UAV-MEC (UAV Edge Computing) architecture has emerged. By offloading computationally intensive tasks to servers at the network edge, it significantly improves the real-time processing capabilities and mission endurance of the swarm.

[0003] However, existing technologies still face significant challenges in addressing collaborative decision-making in complex adversarial environments involving large-scale clusters. On one hand, while multi-agent reinforcement learning collaborative approaches perform well in small-scale clusters, the "curse of dimensionality" is particularly prominent. As the cluster size increases, the joint state-action space grows exponentially, making algorithm convergence difficult. Simultaneously, the dense communication demands between individuals place enormous bandwidth pressure on inter-machine links, severely restricting cluster scalability and real-time decision-making. On the other hand, while layered decoupled collaborative methods reduce complexity to some extent, their inherent fragmented optimization paradigm leads to suboptimal strategies. For example, the upper-layer path planner, when making decisions, cannot foresee the profound impact of path selection on the lower-layer cluster's defense configuration and the computational load of edge nodes. This may lead the cluster to a geometrically optimal but weakly defended or computationally congested region, sacrificing the overall survivability and task efficiency of the cluster. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0005] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides an optimal path selection and defense coordination method for edge computing of unmanned aerial vehicle (UAV) swarms to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: an optimal path selection and defense coordination method for edge computing of unmanned aerial vehicle (UAV) swarms, comprising: A composite mean field driven UAV swarm system model is constructed, wherein an individual state including its position, speed, energy and the amount of computing tasks to be processed is defined for each UAV in the swarm, and a composite mean field is defined to statistically characterize the spatial distribution density, defense resource density and computing load density of the entire swarm. An iterative learning method is used to solve the model to obtain the optimal cooperative strategy. The iterative learning method includes: optimizing the strategy of a single UAV based on a given mean field evolution trajectory to minimize a coupling cost function; and generating new swarm behavior data through large-scale simulation based on the optimized strategy, and updating the mean field evolution trajectory accordingly; repeating the above process until the mean field evolution trajectory converges to the mean field equilibrium state. The optimal collaborative strategy is distributed and executed between the UAV and the edge computing node. The execution process includes: the UAV periodically reporting its own status to its associated edge computing node; the edge computing node aggregating the UAV status within its service area, constructing and predicting the local mean field in real time; and sending the local mean field information to the UAV; the UAV generating and executing collaborative actions including flight control, defense resource allocation, and computing task offloading based on its own status and the received local mean field information.

[0007] As a preferred embodiment of the optimal path selection and defense coordination method for UAV swarm edge computing described in this invention, the composite mean field includes: A location density field is used to characterize the density of UAV distribution at various locations in space at any time. A defense resource field is used to characterize the average intensity of defense resource input formed by drone swarms at various locations in space at any time; And a computational load field, used to characterize the average computational task unloading rate exerted by a UAV at any location in space on its associated edge computing node at any time.

[0008] As a preferred embodiment of the optimal path selection and defense coordination method for UAV swarm edge computing described in this invention, the coupling cost function comprises a weighted sum of at least four of the following terms: One path cost is related to the drone's energy consumption and flight time; A survival cost is positively correlated with the level of external environmental threats and negatively correlated with the value of the defense resource field at the location of the drone; A congestion cost is positively correlated with the value of the location density field at the location of the drone; In addition, a computational cost depends on the drone's computational task offloading decision and the value of the computational load field carried by the target edge computing node.

[0009] As a preferred embodiment of the optimal path selection and defense coordination method for UAV swarm edge computing described in this invention, the method for calculating the survival cost includes: A function value representing external environmental threats is exponentially weakened based on the value of the defense resource density field corresponding to the location of the drone, thereby quantifying the improvement of individual survival probability by cluster collaborative defense.

[0010] As a preferred embodiment of the optimal path selection and defense coordination method for edge computing of UAV swarms described in this invention, the computing cost includes local computing cost and offloading computing cost; wherein, the calculation of the offloading computing cost is positively correlated with the value of the computing load density field of the area served by the edge computing node of the UAV intending to offload the task, so as to reflect the queuing delay caused by the offloading behavior of the swarm computing task.

[0011] As a preferred embodiment of the optimal path selection and defense coordination method for UAV swarm edge computing described in this invention, the solution process of the iterative learning method includes: Under the condition of fixing the mean field evolution trajectory, the problem is transformed into a single agent reinforcement learning problem, and a policy network is trained so that its output action can minimize the coupling cost function. The trained policy network is used to drive the drone swarm to perform simulation, generating new mean field evolution data. This data is then used to train a mean field evolution prediction network to predict the mean field at the next time step based on the mean field at the current time step.

[0012] As a preferred embodiment of the optimal path selection and defense coordination method for UAV swarm edge computing described in this invention, the input of the policy network includes: The individual state of a single drone and the local mean field around the location of the drone, wherein the local mean field is a spatial subset extracted from the global mean field; the output of the policy network is the cooperative action.

[0013] As a preferred embodiment of the optimal path selection and defense coordination method for edge computing of UAV swarms described in this invention, the mean field evolution prediction network adopts an encoder-decoder structure, the input of which is the composite mean field represented in discretized grid form at the current time, and the output is the predicted composite mean field with the same data structure at the next time.

[0014] As a preferred embodiment of the optimal path selection and defense coordination method for UAV swarm edge computing described in this invention, the real-time construction and prediction of the local mean field includes: The local mean field at the current moment is updated using the aggregated UAV state, and a pre-trained mean field evolution prediction network is used to calculate the predicted local mean field for one or more future time steps, and the predicted local mean field is sent to the UAV.

[0015] As a preferred embodiment of the optimal path selection and defense coordination method for UAV swarm edge computing described in this invention, the generation of the coordinated action includes: The individual state and the predicted local mean field information received from the edge computing node are input into a pre-trained policy network deployed locally on the drone for reasoning to obtain the cooperative action.

[0016] Compared with existing technologies, the beneficial effects of this solution are: 1. By employing mean-field game theory, the N-agent intelligent game problem is transformed into an optimization problem between a single agent and the mean field, thus completely decoupling the computational complexity of individual decisions from the cluster size N. This fundamentally overcomes the problem of the state-action space growing exponentially with the cluster size in traditional multi-agent reinforcement learning methods, enabling the proposed solution to be efficiently applied to large-scale cluster scenarios involving hundreds or thousands of drones.

[0017] 2. By constructing a composite average field that includes three dimensions: location, defense, and computation, and designing a coupled cost function that integrates flight costs, survival risks, congestion costs, and computational overhead, UAVs can proactively weigh the construction of cluster defense patterns and the load on edge computing resources when planning their paths. This avoids the decision-making fragmentation and strategy suboptimal problems caused by traditional layered decoupling methods, thereby improving the overall mission success rate and survivability of the cluster.

[0018] 3. Furthermore, this invention leverages the powerful computing capabilities of edge computing nodes to not only construct local mean fields in real time but also utilizes pre-trained evolutionary prediction networks to infer future trends. This enables individual UAVs to make decisions based on predictions of the future (rather than the present) macroscopic environment, thereby avoiding impending congestion areas, seizing advantageous defensive positions on the future battlefield, or selecting edge nodes with lower future computational loads, endowing the cluster with unprecedented intelligent foresight and proactive collaborative capabilities. Simultaneously, this invention constructs an "individual-edge" distributed collaborative architecture, reducing communication bandwidth pressure. The online execution scheme designed in this invention primarily uses a "many-to-one" reporting mode from UAVs to associated edge nodes and a "one-to-many" broadcasting mode from edge nodes to UAVs within the region. This avoids the fully connected or locally dense communication required for state sharing between UAVs in traditional methods, reducing the total amount of data exchange and topology maintenance complexity in the network, and ensuring the real-time performance and robustness of large-scale cluster decision-making in complex electromagnetic environments. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating the overall process of optimal path selection and defense coordination method for edge computing of unmanned aerial vehicle (UAV) swarms according to an embodiment of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0023] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0024] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device 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 the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0025] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0026] Example 1 Reference Figure 1 This is the first embodiment of the present invention, which provides an optimal path selection and defense coordination method for edge computing of unmanned aerial vehicle (UAV) swarms, including: S1. Construct a composite mean field driven UAV swarm system model, wherein an individual state including its position, speed, energy and the amount of computing tasks to be processed is defined for each UAV in the swarm, and a composite mean field is defined to statistically characterize the spatial distribution density, defense resource density and computing load density of the entire swarm.

[0027] It should be noted that this step aims to abstract and transform a physically complex, high-dimensional, large-scale UAV swarm multi-agent interaction problem into a mathematically tractable and highly scalable mean-field game (MFG) model. The specific construction process includes modeling the micro-states and actions of individual UAVs, performing composite mean-field modeling of the swarm's macro-statistical characteristics, and constructing a coupled cost function connecting micro-decision-making with macro-effects.

[0028] Furthermore, a model is created for any representative drone i in the drone swarm, where... N is the total number of drones in the cluster.

[0029] Furthermore, the individual state of drone i at time t is defined. This state is a multi-dimensional vector that comprehensively describes the drone's own situation and serves as the basis for its decision-making. Its specific form is: in, Let be the position coordinate vector of the UAV in three-dimensional space. This is the velocity vector of the drone. The remaining energy of drones is a critical constrained resource. The amount of data or queue length of the computing tasks to be processed on the drone reflects its current computing pressure.

[0030] Furthermore, define the cooperative actions that UAV i can perform at time t. This action involves decision-making across three dimensions: flight, defense, and computation. Its specific form is as follows: in, The flight control input for unmanned aerial vehicles (UAVs) is typically expressed as acceleration or thrust commands, which directly affect their state transitions. and . The strength of defensive resources allocated to a drone, such as electronic jamming power or energy shield output. This decision does not directly change its physical state, but it affects its probability of survival in the environment. Decisions to offload computational tasks. This indicates that the task is processed locally; (in, This indicates that the task is offloaded to the k-th edge computing node, denoted as MEC.

[0031] It should be noted that the core idea of ​​mean-field theory is that when In this context, the complex interactions between a single agent and N-1 other agents can be approximated by its interaction with an "average field" that describes the overall statistical distribution. Based on this, the present invention constructs a composite average field. The value of this composite mean field at spatial point x and time t is composed of the following three subfields: (i) Location density field This field characterizes the distribution density of a drone swarm near spatial point x at time t. Its physical significance lies in quantifying the congestion level and potential collision risk in a specific area. Mathematically, it can be determined by the positions of all drones. We obtained the following through kernel density estimation: in, It is a kernel function, such as a Gaussian kernel, used to smoothly transform discrete UAV positions into a continuous density field.

[0032] (ii) Defensive resource fields This field characterizes the average defensive resource input intensity at spatial point x, formed collaboratively by the entire swarm at time t. Its physical significance lies in quantifying the thickness of the "defensive barrier" formed by the swarm in different regions. Individual drones can obtain better protection by gravitating towards areas with higher field values. Its mathematical form represents the defensive actions of all drones. In its position The sum of density contributions: in, It is a kernel function whose scope of action is related to the characteristics of defensive resources.

[0033] (iii) Calculate the load field This field characterizes the average computational load borne by the edge computing nodes associated with the spatial region x at time t. Unlike the previous two fields, this load field is discretely defined at the locations of K MEC nodes. The computational load of the k-th MEC node at time t. The workload is determined by the total workload contributed by all drones that choose to unload to that node, reflecting the node's activity level. in, It is an indicator function, when The value is 1 if it is true, and 0 otherwise. It is the data rate of the drone's unloading mission.

[0034] Furthermore, by constructing a coupled cost function Micro-decision of individual drones This is closely related to its macroscopic environment (described by the mean field M). Based on this, we define the UAV's objective as selecting a strategy to minimize the expected cumulative value of this cost function over the mission period T. In this embodiment, the cost of this coupled cost function consists of the following four weighted sums: Path cost This reflects the direct costs of flight, primarily related to energy consumption and time. in, and These are weighting coefficients. This represents the instantaneous energy consumption related to thrust.

[0035] Cost of living This item compares the survivability risk of drones with the defensive resource fields of their location. Negative correlation with external environmental threats, and positive correlation with them, demonstrate the value of collaborative defense: in, Given an external threat function, representing the location The inherent threat level (such as the probability of detection by enemy radar). The term indicates the exponential attenuation effect of the defense resources provided by the cluster on the threat. It is a parameter that characterizes the effectiveness of defense resources.

[0036] It should be noted that the survival cost is used to incentivize drones to cluster in areas with high defense resource values, thereby forming effective collective protection.

[0037] Congestion costs This project integrates the drone's decision-making with the location density field. Positive correlation is used to penalize behaviors that enter overcrowded areas, thereby avoiding collision risks and maintaining reasonable cluster formation: in, It is the weighting coefficient.

[0038] Calculation cost This item links the processing cost of the computational task to the computational load. The correlation reflects the scarcity of computing resources and the consequences of task offloading decisions: in, It refers to local processing costs, which are related to the number of tasks pending. The increasing function reflects the energy consumption and time delay of airborne computing. The communication transmission cost (energy consumption, latency) during unloading is related to the distance from the drone to the MEC node. This refers to the queuing delay cost generated at the MEC node, which is related to the node's computational load field. It is a strictly monotonically increasing function (e.g., it can be modeled based on the M / M / 1 queuing model).

[0039] It should be noted that the computational cost enables drones to proactively avoid overloaded MEC nodes when making decisions, thereby achieving load balancing of computing resources at the cluster level.

[0040] S2. An iterative learning method is used to solve the model and obtain the optimal cooperative strategy. The iterative learning method includes: optimizing the strategy of a single UAV based on a given mean field evolution trajectory to minimize a coupling cost function; and generating new cluster behavior data through large-scale simulation based on the optimized strategy, and updating the mean field evolution trajectory accordingly; repeating the above process until the mean field evolution trajectory converges to the mean field equilibrium state.

[0041] Furthermore, after constructing the mean-field game (MFG) model, it is necessary to solve the model during its offline training phase to obtain a strategy that can guide the UAV to make optimal collaborative decisions. Since the MFG model includes a forward individual optimal control problem (based on a given mean field) and a backward macroscopic dynamic evolution problem (driven by individual policies), its analytical solution is often difficult to obtain. Therefore, this invention employs an iterative learning method based on the concept of fictitious play, alternately optimizing the microscopic individual policies and the macroscopic mean-field evolution model, ultimately converging to a stable Nash equilibrium point, namely the mean-field equilibrium (MFE). Based on this, we can decompose an extremely complex N-body game problem into two iteratively solvable subproblems.

[0042] Furthermore, sub-problem one: individual policy optimization under a fixed mean field. That is, under the condition of a fixed mean field evolution trajectory, the problem is transformed into a single-agent reinforcement learning problem, training a policy network.

[0043] Specifically, at the beginning of the k-th iteration, we assume the macroscopic mean-field evolution trajectory of the cluster. It is known and fixed. This evolutionary trajectory can be generated by the mean-field prediction network obtained from the previous iteration. Under this assumption, for any individual drone, the dynamic changes of its environment (including threats, congestion, computational load, etc.) are deterministic and exogenous, no longer affected by the real-time decisions of other individuals. In this way, the originally coupled N-agent game problem is successfully decoupled into a standard single-agent optimal control problem. The goal of this drone is to learn a policy. This makes it follow the strategy and the given mean field During interaction, it can minimize the expected cumulative cost within the task cycle T. : in, It is a discount factor. This represents the coupling cost function defined in step S1.

[0044] Furthermore, to solve the aforementioned optimal control problem, this invention employs a deep reinforcement learning (DRL) method to train a parameterized policy network. .

[0045] Specifically, to ensure the policy has generalization ability and is computationally efficient, the input to the policy network is not the global mean field, but rather the individual states of the UAVs. and its location Local mean field in the surrounding area The combination of . From the global mean field One extracted from Centered on, with dimensions of A spatial subset (or "perceptual slice"). Based on this, the input data for the policy network is... .

[0046] It should be noted that the strategy network input design of the present invention can conform to the physical reality of the limited sensing range of UAV sensors and reduce the dimensionality of the input.

[0047] Furthermore, in this example, the policy network A hybrid architecture is employed to efficiently handle heterogeneous inputs. Its structure is as follows: A multilayer perceptron module containing two hidden layers (128 neurons each, using the ReLU activation function) is used to process individual states in vector form. Extract its features.

[0048] A 3D convolutional neural network (3D-CNN) module containing three convolutional layers (kernel size 3×3×3, stride 1, and channel numbers 16, 32, and 64 respectively) is used to process local mean fields in tensor form. (It includes three channels: location, defense, and computation) to capture the spatial structural characteristics of the macro environment.

[0049] Furthermore, the feature vectors extracted from the two modules are flattened and concatenated, then passed through two fully connected layers (256 neurons each) to finally output the coordinated action. The parameters. For continuous motion. and The policy network will output the mean and standard deviation of a Gaussian distribution; for discrete actions... The policy network will output logical values ​​for the classification distribution.

[0050] Furthermore, in this embodiment, the Proximal Policy Optimization (PPO) algorithm is used to optimize the policy network. Conduct training. During training, provide reward signals. Set to the negative value of the coupling cost function, i.e. Then, the PPO algorithm introduces a clipping mechanism into the objective function to limit the step size of each policy update, ensuring the stability of the training process. The training objective is to maximize the expected cumulative reward (equivalent to minimizing the expected cumulative cost). This is achieved by using a simulation environment (which is based on a given...) The network performs extensive interactive sampling (evolution) to collect trajectory data and update network parameters accordingly. .

[0051] Specifically, in this embodiment, the parameters for training the policy network using the PPO algorithm are set as follows: learning rate. Discount factor PPO trimming parameters The training batch size is 2048 time steps, and the subproblem is considered to have converged when the average cumulative reward no longer increases significantly within an evaluation window (e.g., 100 consecutive rounds).

[0052] Furthermore, sub-problem two: Macroscopic mean-field evolution update based on optimal policy. That is, using the trained policy network to drive the UAV swarm to perform simulation, generating new mean-field evolution data, and using this data to train a mean-field evolution prediction network.

[0053] Specifically, since a fixed mean field was obtained in subproblem one... Optimal individual strategy Next, we need to update the mean field to reflect the true collective behavior when all drones adopt this new strategy. This requires a large-scale swarm simulation: Initialize a cluster containing N drones (e.g., 1000).

[0054] In the simulation environment, all drones are equipped with and execute the current optimal policy network. Make decisions.

[0055] Run the simulation from t=0 to the mission end time T. At each time step, record the position of all drones. Defensive actions and computational unloading decision .

[0056] Based on the definition in step S1, using this full dataset, the strategy is calculated. Driven by this, the true, new mean-field evolution trajectory This process can generate a time series dataset. .

[0057] Furthermore, in order to quickly provide the mean-field evolution trajectory in the next iteration of policy optimization without repeating costly large-scale simulations, we also train a mean-field evolution prediction network. To learn the mapping relationships of macroscopic dynamics .

[0058] Specifically, this prediction network An encoder-decoder structure is adopted; in this example, the U-Net architecture is chosen. U-Net has performance advantages in image-to-image conversion tasks and can also effectively handle the field-to-field prediction problem in this invention. Its structure is as follows: A shrinking path (encoder) gradually extracts abstract features of the mean field at different scales through a series of convolutional and downsampling layers, capturing global contextual information.

[0059] An expansion path (decoder) gradually restores abstract features to their original resolution through a series of upsampling (deconvolution) and convolutional layers, enabling accurate pixel-level (or voxel-level) prediction of the mean field at the next time step.

[0060] Skip connections directly pass the feature maps of the corresponding layers in the encoder to the decoder, enabling the network to combine deep semantic information and shallow detail information, thereby improving prediction accuracy.

[0061] Specifically, the input to the prediction network is the composite mean field at time t, represented in the form of a three-dimensional discretized grid. (Contains three channels). The output is the composite mean field predicted at time t+1. Its output data structure is exactly the same as its input.

[0062] Furthermore, since the encoder is a supervised learning task, it is necessary to utilize the time-series dataset generated in the previous simulation step. As input, As labels, using the Adam optimizer, the learning rate is set to... The batch size is 32. The loss function is achieved by minimizing the mean squared error (MSE) between the predicted and actual fields. To train network parameters : It should be noted that after the training is completed, the results obtained are... It can be based on any initial mean field The autoregressive generation of the mean field evolution trajectory throughout the entire task cycle.

[0063] Furthermore, by repeatedly performing the two sub-problems mentioned above, starting from an initial mean field trajectory... Begin by iteratively generating sequences. In each iteration, the individual policy responds optimally to the macro-environment, which in turn evolves according to the new individual policy. This process continues until the difference between the mean-field evolution trajectories of two consecutive iterations is less than a preset threshold. ,Right now At this point, we consider the model training to have reached a mean-field equilibrium state, and we can then output the policy network. and prediction networks This is the optimal collaborative strategy we are looking for.

[0064] S3. Distribute and execute the optimal coordination strategy between the UAV and the edge computing node. The execution process includes: the UAV periodically reports its own status to its associated edge computing node; the edge computing node aggregates the status of UAVs in its service area, constructs and predicts the local mean field in real time; and sends the local mean field information to the UAV; the UAV generates and executes coordinated actions including flight control, defense resource allocation and computing task offloading based on its own status and the received local mean field information.

[0065] Furthermore, the obtained optimal policy network and mean-field evolution prediction network are deployed to actual UAV swarms and edge computing (MEC) nodes to achieve online, distributed, and real-time collaborative decision-making. Specifically, the global, centralized computational burden is transferred to the network edge, and high-quality swarm collaboration is achieved through the execution process of "individual reporting - edge prediction - individual decision-making" while avoiding direct and dense communication between UAVs.

[0066] It is important to note that the entire execution process occurs at a high frequency (e.g., decision cycles). The distributed control loop runs continuously for 50-100 milliseconds, and the steps are as follows: At the beginning of each decision cycle, time t, each drone i in the cluster will update its latest individual state. The data is reported to the edge computing node currently associated with it via a wireless link. It is important to emphasize that this association can be determined based on dynamic strategies such as proximity in geographic location and strongest signal strength.

[0067] Specifically, the reporting is done via data packets, which contain not only the current status. It also includes information on the defensive and computational actions it performed in the previous cycle, namely... and This historical action information is essential raw data for MEC nodes to construct the defense resource field and computing load field at the current time t.

[0068] It should be noted that by using the aforementioned method of reporting individual drone status, each drone can serve as an information sampling point for the cluster, contributing its micro-state to the MEC. The MEC, by aggregating this discrete micro-information, can reconstruct a local macro-level cluster situation. Compared to the traditional "many-to-many" mode in multi-agent reinforcement learning, this "many-to-one" communication mode of the present invention reduces communication overhead and ensures the scalability of large-scale clusters.

[0069] Furthermore, when edge computing nodes Received within its service area After the drone reports its status, it will perform the following two key tasks: Task 1: Utilizing the aggregated Data from drones is used to calculate the local composite mean field at time t within their service area in real time. This calculation method is consistent with the definition in step S1, but its scope is limited to the local area. The calculation method is as follows: Local location density field : Local defense resource field : Local calculation of load field : in, yes A collection of drones within the service area.

[0070] Task Two: Because it only possesses the mean field information at the current moment, the drone's decisions are reactive. Therefore, to achieve forward-looking decision-making, It will immediately invoke the mean-field evolution prediction network trained in step S2. To predict the local mean field in the future.

[0071] Specifically, the newly constructed local mean field, represented by a three-dimensional grid, will be used. As the initial input to the prediction network, an autoregressive approach is used to generate predicted sequences for the next H time steps. in, This represents the predicted value.

[0072] It should be noted that, through the aforementioned forward-looking prediction of the local mean field in the future time domain, the present invention shifts the complex task of extrapolating the future macroscopic trends of the cluster from the resource-constrained drone end to the edge nodes with stronger computing power. In other words, MEC, through a single forward propagation, can provide all drones within the region with a unified "public forecast" of future environmental evolution. This allows drones to make decisions no longer based on currently outdated information, but on reasonable expectations of the future, thereby proactively avoiding future congestion, collaboratively occupying advantageous defensive positions under future threats, and selecting computing nodes with lower future loads.

[0073] Furthermore, after the above calculations are completed, This will include the predicted local mean field information for the next H time steps. It broadcasts or multicasts to all services within its service range via the downlink. A drone.

[0074] Specifically, the content sent is a serialized three-dimensional tensor in the data packet, which contains a dynamic evolution of the drone's surrounding environment (congestion, defense, computing load) over a future period of time.

[0075] It should be noted that, through the above-described distribution operation, the present invention can feed back macro-level situational awareness and prediction results to micro-level decision-making units. Furthermore, since the distributed information is processed and predicted structured information, rather than the original inter-individual states, the information density is increased, and the corresponding communication efficiency is guaranteed.

[0076] Furthermore, after receiving the predicted mean field information from the MEC, UAV i compares it with its current state. This is then integrated to construct the input vector fed into its local policy network. Specifically, the drone will use the first prediction time step. The mean field information (which is the environmental state most directly related to its next action): Furthermore, the drone will use the aforementioned input vector The policy network is fed into a pre-trained network deployed on the airborne computing unit. In the middle, a fast forward inference calculation is performed, and the following is obtained: It is important to emphasize that, due to policy networks With optimized design and continuously improving computing power of onboard computing units (such as the NVIDIA Jetson series), this inference process can be completed in milliseconds, fully meeting the needs of real-time decision-making.

[0077] Furthermore, the coordinated actions output by the policy network The data is parsed and distributed to the corresponding subsystems of the drone, including: Flight control commands Send the data to the flight control system to adjust the drone's trajectory.

[0078] Deploy defense resource allocation instructions Send to the electronic warfare or energy shield module.

[0079] Decision to offload computing tasks The message is sent to the Communication and Task Management module to determine whether the next computation task should be processed locally or offloaded to MEC node k.

[0080] Furthermore, after performing the aforementioned coordinated actions, the drone's state will change to... Then, the next decision-making cycle begins, and this step is repeated.

[0081] It should be noted that the present invention, by designing a collaborative computing architecture of "drone-edge", can transform the globally optimal strategy trained in step S2 into a scalable, low-latency, and forward-looking distributed online execution scheme, realizing highly collaborative and integrated intelligent decision-making for path, defense, and computing tasks of large-scale drone swarms in complex dynamic environments.

[0082] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0083] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0084] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0085] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0086] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0087] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for optimal path selection and defense coordination of UAV cluster edge computing, characterized in that, include: A composite mean field driven UAV swarm system model is constructed, wherein an individual state including its position, speed, energy and the amount of computing tasks to be processed is defined for each UAV in the swarm, and a composite mean field is defined to statistically characterize the spatial distribution density, defense resource density and computing load density of the entire swarm. An iterative learning method is used to solve the model to obtain the optimal cooperative strategy. The iterative learning method includes: optimizing the strategy of a single UAV based on a given mean field evolution trajectory to minimize a coupling cost function; and generating new swarm behavior data through large-scale simulation based on the optimized strategy, and updating the mean field evolution trajectory accordingly; repeating the above process until the mean field evolution trajectory converges to the mean field equilibrium state. The optimal collaborative strategy is distributed and executed between the UAV and the edge computing node. The execution process includes: the UAV periodically reporting its own status to its associated edge computing node; the edge computing node aggregating the UAV status within its service area, constructing and predicting the local mean field in real time; and sending the local mean field information to the UAV; the UAV generating and executing collaborative actions including flight control, defense resource allocation, and computing task offloading based on its own status and the received local mean field information. 2.The optimal path selection and defense coordination method of UAV cluster edge computing according to claim 1, wherein, The composite average field includes: A location density field is used to characterize the density of UAV distribution at various locations in space at any time. A defense resource field is used to characterize the average intensity of defense resource input formed by drone swarms at various locations in space at any time; And a computational load field, used to characterize the average computational task unloading rate exerted by a UAV at any location in space on its associated edge computing node at any time.

3. The optimal path selection and defense coordination method for UAV swarm edge computing as described in claim 1, characterized in that, The coupling cost function comprises a weighted sum of at least four of the following terms: One path cost is related to the drone's energy consumption and flight time; A survival cost is positively correlated with the level of external environmental threats and negatively correlated with the value of the defense resource field at the location of the drone; A congestion cost is positively correlated with the value of the location density field at the location of the drone; In addition, a computational cost depends on the drone's computational task offloading decision and the value of the computational load field carried by the target edge computing node. 4.The optimal path selection and defense coordination method of UAV cluster edge computing according to claim 3, wherein, The method for calculating the cost of survival includes: The function value representing the threat of the external environment is exponentially weakened based on the value of the defense resource density field corresponding to the location of the drone, thereby quantifying the improvement of the individual survival probability by cluster collaborative defense. 5.The optimal path selection and defense coordination method of UAV cluster edge computing according to claim 3, wherein, The computational cost includes local computational cost and offloading computational cost; wherein, the calculation of the offloading computational cost is positively correlated with the value of the computational load density field of the area served by the edge computing node of the UAV intending to offload the task, in order to reflect the queuing delay caused by the cluster computing task offloading behavior. 6.The optimal path selection and defense coordination method of UAV cluster edge computing according to claim 1, wherein, The solution process of the iterative learning method includes: Under the condition of fixing the mean field evolution trajectory, the problem is transformed into a single agent reinforcement learning problem, and a policy network is trained so that its output action can minimize the coupling cost function. The trained policy network is used to drive the drone swarm to perform simulation, generating new mean field evolution data. This data is then used to train a mean field evolution prediction network to predict the mean field at the next time step based on the mean field at the current time step.

7. The optimal path selection and defense coordination method of UAV cluster edge computing according to claim 6, wherein, The inputs to the policy network include: The individual state of a single drone and the local mean field around the location of the drone, wherein the local mean field is a spatial subset extracted from the global mean field; the output of the policy network is the cooperative action.

8. The optimal path selection and defense coordination method for UAV swarm edge computing as described in claim 6, characterized in that, The mean field evolution prediction network adopts an encoder-decoder structure. Its input is the composite mean field represented in discretized raster form at the current time, and its output is the predicted composite mean field with the same data structure at the next time. 9.The optimal path selection and defense coordination method of UAV cluster edge computing according to claim 1, wherein, The real-time construction and prediction of the local mean field includes: The local mean field at the current moment is updated using the aggregated UAV state, and a pre-trained mean field evolution prediction network is used to calculate the predicted local mean field for one or more future time steps, and the predicted local mean field is sent to the UAV.

10. The optimal path selection and defense coordination method for UAV swarm edge computing as described in claim 1 or 9, characterized in that, Generating the cooperative action includes: The individual state and the predicted local mean field information received from the edge computing node are input into a pre-trained policy network deployed locally on the drone for reasoning to obtain the cooperative action.

Images