Layered decision-making method and system for end side cloud of unmanned aerial vehicle agent

Abstract

The invention discloses an unmanned aerial vehicle agent end side cloud hierarchical decision-making method and system, and the method comprises the steps: S1, deploying a compressed lightweight deep learning model at an unmanned aerial vehicle terminal, and executing a real-time preliminary decision of a simple or emergency task; s2, constructing a dynamic task unloading mechanism, and unloading the complex task from the unmanned aerial vehicle terminal to an edge computing node; s3, carrying out preprocessing and local decision making on the task by the edge computing node, and further judging whether the task needs to be forwarded to the cloud data center or not; s4, the cloud data center carries out global optimization decision making for calculation-intensive or data-intensive tasks; and S5, monitoring network and computing node states in real time, and dynamically adjusting end, edge and cloud computing resource allocation to realize dynamic balance of decision precision and real-time response. According to the technical scheme disclosed by the invention, the computing power and energy consumption bottleneck of the unmanned aerial vehicle terminal can be effectively relieved, the computing resource allocation is dynamically optimized, and the real-time response capability and decision-making precision of complex task processing of the unmanned aerial vehicle are improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and unmanned aerial vehicle (UAV) technology, and more particularly, to a hierarchical decision-making method and system for UAV intelligent agents at the edge and in the cloud. Background Art

[0002] In recent years, with the rapid development of UAV technology, UAVs have been widely used in many fields such as agricultural monitoring, environmental inspection, target tracking, search and rescue operations, etc. However, in the actual operation of UAVs, due to the large limitations in computing power and energy consumption of UAV terminals, it is difficult for them to independently meet the real-time processing requirements of complex tasks.

[0003] Currently, in order to solve the real-time processing requirements of UAV complex tasks, an edge-cloud collaborative computing architecture has been proposed in the prior art. The edge-cloud collaborative computing architecture is a distributed computing framework that organically combines terminal devices, edge computing nodes, and cloud computing centers, aiming to meet the diverse scenario requirements such as low latency, high reliability, and large bandwidth through resource collaboration, task scheduling, and data processing optimization. When a UAV executes a complex task, the task is reasonably allocated to the UAV terminal, edge computing node, and cloud data center to achieve efficient processing. However, the current technical solutions using the edge-cloud collaborative computing architecture often have difficulty in flexibly adapting to changes in the network environment and fluctuations in the computing load, lacking the ability of dynamic decision-making and resource scheduling, which limits the real-time performance and decision-making accuracy of UAV task execution.

[0004] Therefore, how to provide a hierarchical decision-making method and system for UAV intelligent agents at the edge and in the cloud, which can effectively alleviate the computing power and energy consumption bottlenecks of UAV terminals, dynamically optimize the allocation of computing resources, and improve the real-time response ability and decision-making accuracy of UAV complex task processing, has become a technical problem that needs to be solved urgently by those skilled in the art. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a hierarchical decision-making method and system for UAV intelligent agents at the edge and in the cloud, which can effectively alleviate the computing power and energy consumption bottlenecks of UAV terminals, dynamically optimize the allocation of computing resources, and improve the real-time response ability and decision-making accuracy of UAV complex task processing.

[0006] The technical solutions provided by the present invention are as follows: The present invention provides a method for hierarchical decision-making of an intelligent agent at the edge and cloud of an unmanned aerial vehicle (UAV), including the following steps: S1. Real-time preliminary decision-making at the terminal: Construct a lightweight deep learning model, deploy the compressed lightweight deep learning model to the UAV terminal, and the terminal makes a quick preliminary decision on simple or urgent tasks based on the compressed lightweight deep learning model; S2. Dynamic task offloading at the terminal: Build a dynamic task offloading mechanism in the UAV terminal, model the resource requirements of the current task according to task complexity, data volume, and network status, and combine the self-attention mechanism to adaptively offload some computing tasks from the terminal to the edge computing node; S3. Task processing at the edge computing node: The edge computing node obtains a local decision-making model through knowledge distillation, processes the tasks offloaded from the terminal, and further determines whether the tasks need to be forwarded to the cloud data center; S4. Global optimization decision-making at the cloud: Forward the tasks determined by the edge computing node as computationally intensive or data-intensive to the cloud data center, and the cloud data center makes a global optimization decision based on a large-scale pre-trained model; S5. Dynamic load awareness and resource scheduling: Real-time monitor and schedule the computing resources of tasks through the dynamic load awareness mechanism among the terminal, edge, and cloud layers, optimize the overall allocation of resources, and dynamically balance decision-making accuracy and real-time response.

[0007] Further, in a preferred embodiment of the present invention, in step S1, the method for compressing the lightweight deep learning model includes, but is not limited to, one or more combinations of network pruning, low-rank decomposition, and quantization-aware training.

[0008] Further, in a preferred embodiment of the present invention, in step S1, the steps of compressing the lightweight deep learning model specifically include: S101. Train the original deep learning teacher model in the cloud data center to obtain initial network parameters; S102. Prune redundant parameters of the teacher model using network pruning; redundant parameter pruning includes calculating the L1 norm of the network weights: , where represents the value of the -th weight in the initial network. Subsequently, set a threshold , and set the weights with importance lower than the threshold to zero to determine the importance of the weights; S103. Perform low-rank decomposition on the model after network pruning, and decompose the model weight matrix through singular value decomposition. The decomposition formula is: ; where, is the left singular matrix retaining the first columns, is the diagonal matrix containing the first A diagonal matrix of singular values, is the right singular matrix that retains the first columns; S104. Perform quantization-aware training on the model after low-rank decomposition compression. Map the floating-point parameters to the integer space through the quantization function to complete model compression. The implementation formula is as follows: ; where, is the quantization scale factor, is the quantization zero point.

[0009] Furthermore, in a preferred embodiment of the present invention, in step S2, the method for adaptively offloading part of the computing tasks from the terminal to the edge computing node is as follows: Generate the attention weights of the task features based on the self-attention mechanism, evaluate the urgency and processing priority of task offloading, thereby dividing the dynamic computing loads on the terminal side and the edge side, and perform offloading decisions according to the dynamic computing loads to offload part of the computing tasks from the terminal to the edge computing node. It specifically includes the following steps: S201. Task feature extraction and encoding: Extract the feature vector of the task received by the UAV terminal, denoted as the task feature vector , and the task feature vector includes the data volume size, computing complexity, network status, and urgency of the task. Subsequently, perform task feature embedding on the task feature vector through linear mapping: ; where, and are the learnable parameter matrix and bias term; S202. Self-attention mechanism to calculate attention weights: Use the scaled dot-product attention mechanism in the self-attention mechanism to process the task feature embedding, calculate the attention weights of the current task features, and obtain the attention vector: ; where represents the current task feature, and represent the set of historical task features, represents the dimension of the feature vector; S203. Calculate the task offloading urgency and priority: According to the attention vector, and combined with the non-linear activation function, evaluate the urgency of the task and the processing priority : ; S204. Dynamic task offloading decision: Based on the urgency and processing priority to determine the strategy and proportion of task offloading. Subsequently, based on the strategy and proportion of task offloading, divide the dynamic computing load between the edge side and the terminal side, and execute the task offloading decision according to the dynamic computing load; S205. Dynamic adjustment of task offloading ratio: Preset the task offloading ratio as , and then determine the task offloading ratio based on the urgency as follows: ; When the task urgency is high, i.e., , then , indicating that the task is all locally executed; When the task urgency is low, i.e., , indicating that most or all of the tasks are offloaded.

[0010] Furthermore, in a preferred embodiment of the present invention, in step S204, the steps of executing the task offloading decision specifically include: Set thresholds , including a priority threshold , a complexity threshold , and a data volume threshold ; If the task offloading urgency exceeds the priority threshold , it indicates that the task is urgent and needs to be immediately locally processed without task offloading; If the task priority is low, the complexity is high, or the data volume exceeds the data volume threshold , then offload the task to the edge computing node; The specific decision is: ; Furthermore, in a preferred embodiment of the present invention, in step S3, the method for obtaining the local decision model is: Based on knowledge distillation, pre-train a high-precision teacher model in the cloud data center, and then use the edge computing node to train a lightweight student model under the guidance of the teacher model, so as to quickly complete task preprocessing and local decision-making on the edge side.

[0011] Furthermore, in a preferred embodiment of the present invention, in step S3, the steps of obtaining the local decision model specifically include: S301. Train the teacher model: Train a high-precision teacher model in the cloud data center ; Among them, is the input task data, is the corresponding true label, is the total number of categories, is the prediction probability of the teacher model for the th category; S302, Obtain soft labels: Adjust the logit vector output by the teacher model for temperature to generate a soft label probability distribution: ; Among them, is the logit output of the teacher model, which is the score before softmax; is the soft label after temperature processing; is the temperature parameter; S303, Train the student model: Deploy the lightweight student model at the edge computing node, and construct a joint distillation loss function according to the soft label provided by the teacher model as the guiding information: ; Among them, is the cross-entropy loss of the student model based on the true label, is the KL divergence between the student and the teacher, , are the softmax outputs of the student and the teacher respectively after using the same temperature parameter, weight coefficient, is the temperature square term; S304, Model compression and edge deployment: Further perform model compression operations on the student model and deploy it in the edge computing node to perform fast preprocessing and local decision-making on the tasks unloaded by the terminal.

[0012] Furthermore, in a preferred manner of the present invention, in step S4, the large-scale pre-trained model adopts a Transformer structure and is pre-trained using the large-scale environmental data accumulated in the cloud; the steps of performing global optimization decisions based on the large-scale pre-trained model specifically include: S401, Construct the input data format: Structurally process the complex task data uploaded by the drone to construct an input sequence , expressed as: ; Among them, represents the input task feature vector, is the position encoding; S402. Transformer Encoding Layer Processing: The input sequence is processed through multiple layers by a Transformer encoder. Each layer includes a multi-head self-attention mechanism and a feed-forward network. Attention calculation: ; Subsequently, after parallel attention calculations are performed on multiple heads, splicing processing is carried out: ; Then, a two-layer linear transformation and activation function processing are performed on the output vector at each position through a feed-forward network: ; S403. Design of Pre-training Tasks: The model is pre-trained based on a large amount of environmental data accumulated in the cloud. The training objectives include, but are not limited to, path planning loss, image recognition loss, and delay prediction loss. The total loss function is expressed as: ; Among them, , , represent the path optimization loss, image recognition accuracy loss, and delay prediction error loss respectively; S404. Inference Phase Decision Output: In the inference phase, based on the pre-trained Transformer model, global optimization decisions are made for computationally intensive or data-intensive tasks forwarded by edge computing nodes. The outputs include, but are not limited to, path prediction results, target recognition categories, or control strategies.

[0013] Furthermore, in a preferred embodiment of the present invention, in step S5, the method for optimizing the overall resource allocation is as follows: The network bandwidth, delay, and load status of each computing node are monitored in real time according to the dynamic load awareness mechanism. Subsequently, a multi-objective optimization algorithm is used to dynamically adjust the computing resource allocation ratio among the terminal, edge, and cloud, specifically including the following steps: S501. Construction of the State Space: Construct a system state vector , and monitor the network bandwidth , communication delay , and the resource utilization rates of the three types of computing nodes, namely the terminal, edge, and cloud, which are expressed as: ; Among them, , , represent the CPU / GPU resource utilization rates on the terminal, edge, and cloud sides respectively; S502. Definition of the Action Space: Define a resource scheduling action space , which represents the computing resource allocation ratios of the three types of nodes, namely the terminal, edge, and cloud, and satisfies: ; Among them, , , are the computing resource ratios allocated to the terminal, edge, and cloud respectively; S503. Design the reward function: Using the decision delay , energy consumption and task processing accuracy as the optimization objectives, construct a multi-objective optimization reward function : ; Among them, is the weighting coefficient; S504. Reinforcement learning decision mechanism: Adopt a resource scheduling strategy based on reinforcement learning to optimize the reward function: ; Among them, is the discount factor, is the policy function; S505. Execute resource scheduling: According to the resource allocation ratio output by the policy function, control the distribution and scheduling of computing tasks among the terminal, edge, and cloud layers, and dynamically optimize the allocation of computing resources.

[0014] In addition, this application also provides another technical solution: This application discloses a hierarchical decision-making system for the drone intelligent agent's edge and cloud. The hierarchical decision-making system for the drone intelligent agent's edge and cloud includes: a drone terminal module: built-in with a lightweight deep learning model, a model compression module, and a dynamic task offloading module based on the self-attention mechanism, used to process emergency or simple tasks in real time and adaptively offload complex tasks according to task characteristics; an edge computing node module: used to deploy a lightweight local decision-making model obtained through knowledge distillation, preprocess and make local decisions on the tasks offloaded by the terminal, and determine whether the tasks are further forwarded to the cloud; a cloud data center module: configured with a large-scale pre-trained model based on the Transformer structure, used to execute global optimization decisions for the compute-intensive or data-intensive tasks forwarded by the edge computing node; a dynamic load perception and resource scheduling module: used to monitor the network and computing load status in real time, and realize dynamic scheduling and optimized allocation of computing resources among the drone terminal, edge computing node, and cloud data center based on a multi-objective optimization algorithm.

[0015] An edge-cloud hierarchical decision-making method and system for an unmanned aerial vehicle intelligent agent provided by the present invention, wherein the edge-cloud hierarchical decision-making method comprises the following steps: S1. Terminal real-time preliminary decision-making: construct a lightweight deep learning model, deploy the compressed lightweight deep learning model to the unmanned aerial vehicle terminal after compression processing, and the terminal performs rapid preliminary decision-making on simple or urgent tasks based on the compressed lightweight deep learning model; S2. Terminal dynamic task offloading: construct a dynamic task offloading mechanism in the unmanned aerial vehicle terminal, model the resource requirements of the current task according to task complexity, data volume and network status, and combine the self-attention mechanism to adaptively offload part of the computing tasks from the terminal to the edge computing node; S3. Edge computing node task processing: the edge computing node obtains a local decision-making model through knowledge distillation, processes the tasks offloaded from the terminal, and further determines whether the tasks need to be forwarded to the cloud data center; S4. Cloud global optimization decision-making: forward the tasks determined by the edge computing node as computationally intensive or data-intensive tasks to the cloud data center, and the cloud data center makes global optimization decisions based on a large-scale pre-trained model; S5. Dynamic load awareness and resource scheduling: monitor and schedule the computing resources of tasks in real time through the dynamic load awareness mechanism among the terminal, edge and cloud layers, optimize the overall allocation of resources, and dynamically balance decision-making accuracy and real-time response.For the UAV intelligent agent edge-cloud hierarchical decision-making method disclosed in the present invention, the decision-making method is divided into three layers, specifically the three-layer structure of "UAV terminal - edge computing node - cloud". The tasks are hierarchically processed according to the complexity of the computing tasks and the requirements of real-time performance, so as to achieve efficient, real-time, and intelligent task decision-making under limited computing power. Among them, the edge-cloud hierarchical decision-making method mainly includes five major steps: The first step is real-time preliminary decision-making at the terminal, that is, using a compressed lightweight deep learning model on the UAV terminal, and using the lightweight deep learning model to perform rapid preliminary decision-making on simple or urgent tasks to meet the real-time requirements. The second step is task offloading at the terminal. To meet the real-time processing requirements of complex tasks, some tasks received by the UAV terminal are offloaded to the edge computing node for processing. In this step, a dynamic task offloading mechanism in the UAV terminal is constructed, and the resource requirements of the tasks are modeled through a self-attention mechanism, and the tasks are offloaded to the edge computing node to enable task preprocessing and local decision-making. The third step is task processing at the edge computing node, that is, using knowledge distillation technology to obtain a local decision-making model, and the edge computing node uses the local decision-making model to process the offloaded tasks and further determines whether the tasks need to be forwarded to the cloud data center. The fourth step is global optimization decision-making at the cloud. The cloud receives computationally intensive or data-intensive tasks forwarded from the edge computing node. For computationally intensive or data-intensive tasks, a large-scale pre-trained model based on Transformer is used for global optimization decision-making, which can improve the decision-making accuracy of complex task processing. The last step is dynamic load awareness and resource scheduling, that is, by real-time monitoring of network bandwidth, latency, and the load status of each computing node, dynamically adjusting the allocation ratio of computing resources between the terminal, edge, and cloud, optimizing the allocation of computing resources, and achieving a dynamic balance between the accuracy and real-time response of task decision-making, being able to flexibly adapt to changes in the network environment and fluctuations in computing load, and improving the ability of dynamic decision-making and resource scheduling. It can be seen that the technical solution involved in the present invention, compared with the prior art, can effectively alleviate the computing power and energy consumption bottlenecks of the UAV terminal, dynamically optimize the allocation of computing resources, improve the real-time response ability and decision-making accuracy of UAV complex task processing, and can be widely applied to scenarios such as UAV task planning, target tracking, obstacle avoidance, and inspection and monitoring, with significant practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0017] Figure 1It is a flowchart of the steps of the method for hierarchical decision-making of the drone intelligent agent edge cloud involved in the embodiments of the present invention; Figure 2 It is a flowchart of the steps of the method for compressing and processing the lightweight deep learning model involved in the embodiments of the present invention; Figure 3 It is a flowchart of the steps of adaptively offloading part of the computing tasks from the terminal to the edge computing node involved in the embodiments of the present invention; Figure 4 It is a flowchart of the steps of obtaining the local decision-making model involved in the embodiments of the present invention; Figure 5 It is a flowchart of the steps of performing global optimization decision-making based on the large-scale pre-trained model involved in the embodiments of the present invention; Figure 6 It is a flowchart of the steps of optimizing the overall allocation of resources involved in the embodiments of the present invention. Detailed implementation manners

[0018] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present invention.

[0019] It should be noted that when an element is referred to as "fixed to" or "disposed on" another element, it can be directly on the other element or indirectly disposed on the other element; when an element is referred to as "connected to" another element, it can be directly connected to the other element or indirectly connected to the other element.

[0020] It should be understood that the orientation or positional relationship indicated by the terms "length", "width", "upper", "lower", "front", "rear", "first", "second", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation to the present invention.

[0021] In addition, the terms "first" and "second" are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality of" or "several" means two or more, unless otherwise specifically defined.

[0022] It should be noted that the structures, ratios, sizes, etc. shown in the drawings of this specification are only used to cooperate with the content disclosed in the specification for those skilled in the art to understand and read, and are not used to limit the conditions for the implementation of the present invention. Therefore, they do not have any technical substance. Any modification of the structure, change in the proportional relationship, or adjustment of the size, without affecting the efficacy and purpose achievable by the present invention, should still fall within the scope covered by the technical content disclosed in the present invention.

[0023] Please refer to Figures 1 to 6As shown in the figure, the present invention provides a method and system for hierarchical decision-making of the drone intelligent agent edge-cloud, which integrates model compression, self-attention mechanism, knowledge distillation, Transformer pre-trained model, and dynamic load perception mechanism, and can optimize the computing resource scheduling among the drone terminal, edge computing nodes, and cloud data center. In the method for hierarchical decision-making of the drone intelligent agent edge-cloud, the intelligent decision-making of the drone is divided into a three-layer structure of "terminal, edge, and cloud", and the computing tasks are hierarchically processed according to the task complexity and real-time requirements, so as to achieve efficient, real-time, and intelligent task decision-making under limited computing power. The specific steps are as follows: S1. Terminal real-time preliminary decision-making: Construct a lightweight deep learning model, deploy the compressed lightweight deep learning model to the drone terminal, and the terminal makes a quick preliminary decision on simple or urgent tasks based on the compressed lightweight deep learning model; S2. Terminal dynamic task offloading: Build a dynamic task offloading mechanism in the drone terminal, model the resource requirements of the current task according to the task complexity, data volume, and network status, and adaptively offload some computing tasks from the terminal to the edge computing nodes by combining the self-attention mechanism; S3. Edge computing node task processing: The edge computing node obtains a local decision-making model through knowledge distillation, processes the tasks offloaded from the terminal, and further determines whether the tasks need to be forwarded to the cloud data center; S4. Cloud global optimization decision-making: Forward the tasks determined by the edge computing node to be computationally intensive or data-intensive to the cloud data center, and the cloud data center makes a global optimization decision based on the large-scale pre-trained model; S5. Dynamic load perception and resource scheduling: Real-time monitor and schedule the computing resources of tasks through the dynamic load perception mechanism among the three layers of the terminal, edge, and cloud, optimize the overall resource allocation, and dynamically balance the decision-making accuracy and real-time response. The technical solution involved in the present invention can effectively alleviate the computing power and energy consumption bottlenecks of the drone terminal, dynamically optimize the computing resource allocation, and improve the real-time response ability and decision-making accuracy of the drone for complex task processing.

[0024] The following specifically elaborates on the method for hierarchical decision-making of the drone intelligent agent edge-cloud disclosed in the present invention in combination with specific embodiments. The method for hierarchical decision-making of the drone intelligent agent edge-cloud specifically includes the following steps: S1. Terminal real-time preliminary decision-making: Construct a lightweight deep learning model, deploy the compressed lightweight deep learning model to the drone terminal, and the terminal makes a quick preliminary decision on simple or urgent tasks based on the compressed lightweight deep learning model.

[0025] Among them, in the specific embodiments of the present invention, step S1 is used to implement real-time decision-making task processing under the computing power and energy consumption limitations of the terminal device; in this step, a lightweight deep learning model is used on the UAV terminal, and then, using model compression technology, the lightweight deep learning model is compressed and deployed on the UAV terminal. The lightweight deep learning model is used to achieve rapid preliminary decision-making for simple or complex tasks to meet the real-time requirements.

[0026] Specifically, in the specific embodiments of the present invention, in step S1, the method for compressing the lightweight deep learning model includes, but is not limited to, one or a combination of more of network pruning, low-rank decomposition, and quantization-aware training.

[0027] Specifically, in step S1, after the lightweight deep learning model is constructed on the UAV terminal, the model needs to be compressed. The purpose of model compression is to optimize the model structure and parameters to significantly reduce resource consumption and improve efficiency while maintaining high performance. After compression, the model volume can be significantly reduced, the inference speed can be accelerated, the energy consumption and cost can be reduced, and the performance and efficiency can be balanced; in the embodiments of the present invention, the method for compressing the lightweight deep learning model includes one or a combination of more of network pruning, low-rank decomposition, and quantization-aware training.

[0028] Specifically, in the specific embodiments of the present invention, in step S1, the steps for compressing the lightweight deep learning model specifically include: S101. Train the original deep learning teacher model in the cloud data center to obtain initial network parameters; S102. Use network pruning to remove redundant parameters of the teacher model; the removal of redundant parameters includes calculating the L1 norm of the network weights: ; Among them, represents the value of the -th weight in the initial network. Subsequently, a threshold is set, and weights with importance lower than the threshold , weights with importance lower than the threshold are set to zero to determine the importance of the weights; S103. Perform low-rank decomposition on the model after network pruning, and decompose the model weight matrix through singular value decomposition. The decomposition formula is: ; Among them, is the left singular matrix retaining the first columns, is the diagonal matrix containing the first singular values, is the right singular matrix retaining the first S104, implement quantization-aware training on the model after low-rank decomposition and compression, through the quantization function Map the floating-point parameters to the integer space to complete the model compression. The implementation formula is as follows: ; in, is the quantitative scale factor, is the quantized zero point.

[0029] In an embodiment of the present invention, step S1 may use one or more combinations of network pruning, low-rank decomposition, and quantization-aware training techniques for model compression; wherein the network pruning is used to remove redundant parameters, retain key structures, and make the model parameters tend to be sparse through L1 norm regularization, that is, many parameters become zero, simplify the model structure and automatically perform feature selection, remove redundant features, and achieve the technical effect of compressing the model; secondly, for the low-rank decomposition, the weight matrix of the model is decomposed into a combination of low-rank matrices to reduce parameter data, thereby achieving model compression. In step S1, the model weight matrix is ​​decomposed by singular value decomposition, and the decomposition formula is used , Before retention The left singular matrix of columns, Is included before The diagonal matrix of singular values, Before retention The right singular matrix of the columns; keep the previous The largest singular values, and retain the corresponding matrix , , Before The quantization-aware training is a technology that simulates quantization operations during model training, which can convert the model from high precision to low precision. The quantization-aware training achieves compression by reducing the numerical precision of model parameters, and utilizes adaptive optimization in the training phase so that the low-precision model can still maintain accuracy close to the original model after compression, avoiding direct quantization that may cause the model to cause performance collapse. The lightweight deep learning model compressed by the above steps is deployed on the drone terminal to realize real-time decision-making task processing of the terminal device under the constraints of computing power and energy consumption.

[0030] S2. Dynamic task offloading at the terminal: A dynamic task offloading mechanism is built in the drone terminal. According to the task complexity, data volume and network status, the resource requirements of the current task are modeled in combination with the self-attention mechanism, and some computing tasks are adaptively offloaded from the terminal to the edge computing node.

[0031] Among them, in the embodiment of the present invention, step S2 is terminal task offloading. To meet the real-time processing requirements of complex tasks, some tasks received by the UAV terminal are offloaded to the edge computing node for processing. In this step, a dynamic task offloading mechanism in the UAV terminal is utilized, and the resource requirements of the tasks are modeled through the self-attention mechanism, and the tasks are offloaded to the edge computing node to enable task preprocessing and local decision-making.

[0032] Specifically, in a specific embodiment of the present invention, in step S2, the method of adaptively offloading some computing tasks from the terminal to the edge computing node is as follows: generating the attention weights of task features based on the self-attention mechanism, evaluating the urgency and processing priority of task offloading, thereby dividing the dynamic computing load between the terminal side and the edge side, and making an offloading decision according to the dynamic computing load to offload some computing tasks from the terminal to the edge computing node.

[0033] Among them, in the embodiment of the present invention, the offloading basis for tasks from the UAV terminal to the edge computing node is divided by the dynamic computing load. The self-attention mechanism in step S2 generates the attention weights of task features, which is used to evaluate the urgency and processing priority of task offloading, so as to realize the division of the dynamic computing load between the terminal side and the edge side. Specifically, in a specific embodiment of the present invention, the steps of adaptively offloading some computing tasks from the terminal to the edge computing node specifically include: S201. Task feature extraction and encoding: Extract the feature vector of the task received by the UAV terminal, denoted as the task feature vector. The task feature vector includes the data volume size, computational complexity, network status, and urgency of the task. Subsequently, task feature embedding is performed on the task feature vector through linear mapping: ; Among them, and are learnable parameter matrices and bias terms; S202. Self-attention mechanism calculates attention weights: Use the scaled dot product attention mechanism in the self-attention mechanism to process the task feature embedding, calculate the attention weights of the current task feature, and obtain the attention vector: ; Among them represents the current task feature, and represent the set of historical task features, represents the dimension of the feature vector; S203. Calculate the task offloading urgency and priority: According to the attention vector, and combined with the non-linear activation function, evaluate the urgency and processing priority : ; S204. Dynamic task offloading decision: Based on the urgency and processing priority to determine the strategy and ratio of task offloading. Subsequently, based on the strategy and ratio of task offloading, divide the dynamic computing load between the edge side and the terminal side, and execute the task offloading decision according to the dynamic computing load; S205. Dynamic adjustment of task offloading ratio: Preset the task offloading ratio as , and then based on the urgency to determine the task offloading ratio : ; When the task urgency is high, i.e., , then , indicating that the task is all executed locally; When the task urgency is low, i.e., , indicating that most or all of the tasks are offloaded.

[0034] Specifically, in a specific embodiment of the present invention, in step S204, the steps of executing the task offloading decision specifically include: setting thresholds , including a priority threshold , a complexity threshold , and a data volume threshold ; If the task offloading urgency exceeds the priority threshold , it indicates that the task is urgent and needs to be processed locally immediately without task offloading; If the task priority is low, the complexity is high, or the data volume exceeds the data volume threshold , then offload the task to the edge computing node; The specific decision is: ;

[0035] S3. Task processing by the edge computing node: The edge computing node obtains a local decision model through knowledge distillation, processes the tasks offloaded from the terminal, and further determines whether the tasks need to be forwarded to the cloud data center.

[0036] Among them, in an embodiment of the present invention, in step S3, the task processing by the edge computing node, that is, using the knowledge distillation technology to obtain a local decision model, and the edge computing node uses the local decision model to process the offloaded tasks and further determines whether the tasks need to be forwarded to the cloud data center.

[0037] Specifically, in a specific embodiment of the present invention, in step S3, the method for obtaining the local decision-making model is as follows: Based on knowledge distillation, a high-precision teacher model is pre-trained in the cloud data center, and then an edge computing node is used to guide the training of a lightweight student model with the teacher model, so as to quickly complete task preprocessing and local decision-making on the edge side.

[0038] Specifically, in a specific embodiment of the present invention, in step S3, the steps of obtaining the local decision-making model specifically include: S301. Training the teacher model: Training a high-precision teacher model in the cloud data center , to minimize the standard classification or regression loss: ; Wherein, is the input task data, is the corresponding true label, is the total number of categories, is the prediction probability of the teacher model for the th category; S302. Obtaining the soft label: Adjust the temperature of the logit vector output by the teacher model to generate a soft label probability distribution: ; Wherein, is the logit output of the teacher model, which is the score before softmax; is the soft label after temperature processing; is the temperature parameter, and the temperature parameter can be used to smooth the probability distribution; S303. Training the student model: Deploying a lightweight student model in the edge computing node, and constructing a joint distillation loss function according to the soft label provided by the teacher model as the guiding information: ; Wherein, is the cross-entropy loss of the student model based on the true label, is the KL divergence between the student and the teacher, , are the softmax outputs of the student and the teacher after using the same temperature parameter respectively, is the weight coefficient, and the weight coefficient is used to balance the two parts of the loss, is the temperature square term, and setting the temperature square term can perform gradient adjustment in backpropagation; S304. Model compression and edge deployment: Further perform model compression operations on the student model and deploy it in the edge computing node. The compressed student model is used to quickly preprocess and locally decide on the tasks unloaded by the terminal.

[0039] S4, Cloud Global Optimization Decision: Forward tasks determined by the edge computing nodes as computationally intensive or data-intensive to the cloud data center, and the cloud data center makes global optimization decisions based on large-scale pre-trained models.

[0040] Specifically, in a specific embodiment of the present invention, in step S4, the large-scale pre-trained model adopts a Transformer structure and is pre-trained using the large-scale environmental data accumulated in the cloud.

[0041] In a specific embodiment of the present invention, to improve the decision accuracy for complex task processing, in step S4, the cloud data center uses a large-scale pre-trained model based on Transformer to make global optimization decisions for computationally intensive or data-intensive tasks.

[0042] Specifically, in a specific embodiment of the present invention, the steps of making global optimization decisions based on the large-scale pre-trained model specifically include: S401, Constructing the input data format: Structurally process the complex task data uploaded by the drone to construct an input sequence , expressed as: ; Among them, represents the input task feature vector, is the positional encoding; S402, Transformer Encoding Layer Processing: Perform multi-layer processing on the input sequence through the Transformer encoder. Each layer includes a multi-head self-attention mechanism and a feed-forward network. Attention calculation: ; Subsequently, perform splicing processing after parallel execution of attention calculations for multiple heads: ; Then, perform two-layer linear transformation and activation function processing on the output vector of each position through the feed-forward network: ; S403, Designing the Pre-training Task: Pre-train the model based on the large-scale environmental data accumulated in the cloud. The training objectives include but are not limited to path planning loss, image recognition loss, and delay prediction loss. The total loss function is expressed as: ; Among them, , , respectively represent the path optimization loss, the image recognition accuracy loss, and the delay prediction error loss; S404, Decision output in the inference stage: In the inference stage, based on the pre-trained Transformer model, global optimization decisions are made on the computationally intensive or data-intensive tasks forwarded by the edge computing nodes, and the outputs include but are not limited to path prediction results, target recognition categories, or control strategies.

[0043] S5, Dynamic load awareness and resource scheduling: Through the dynamic load awareness mechanism among the terminal, edge, and cloud layers, the computing resources of tasks are monitored and scheduled in real time to optimize the overall resource allocation and dynamically balance the decision accuracy and real-time response.

[0044] Specifically, in the specific embodiment of the present invention, in step S5, the method for optimizing the overall resource allocation is as follows: According to the dynamic load awareness mechanism, the network bandwidth, delay, and the load status of each computing node are monitored in real time, and then a multi-objective optimization algorithm is used to dynamically adjust the computing resource allocation ratio among the terminal, edge, and cloud.

[0045] Among them, the multi-objective optimization algorithm in step S5 includes reinforcement learning or multi-armed bandit algorithm.

[0046] Specifically, in the specific embodiment of the present invention, the steps of optimizing the overall resource allocation specifically include: S501, Constructing the state space: Constructing the system state vector , monitoring the network bandwidth in real time , communication delay , and the resource utilization rates of the three types of computing nodes of the terminal, edge, and cloud, which are expressed as: ; Among them, , , respectively represent the CPU / GPU resource utilization rates on the terminal, edge, and cloud sides; S502, Defining the action space: Defining the resource scheduling action space , to represent the computing resource allocation ratios of the three types of nodes of the terminal, edge, and cloud, and making them satisfy: ; Among them, , , are the computing resource ratios allocated to the terminal, edge, and cloud respectively; S503, Designing the reward function, the reward function is used for reinforcement learning: Taking the decision delay , energy consumption and task processing accuracy as the optimization objectives, constructing a multi-objective optimization reward function : ; Among them, is the weighting coefficient; S504, Reinforcement learning decision-making mechanism: Adopt a resource scheduling strategy based on reinforcement learning to optimize the reward function: ; Among them, is the discount factor, is the policy function; S505, Execute resource scheduling: According to the resource allocation ratio output by the policy function, control the distribution and scheduling of computing tasks among the terminal, edge, and cloud layers, and dynamically optimize the allocation of computing resources.

[0047] Specifically, step S504 is used to optimize the reward function, which is optimized through the policy function; in the embodiments of the present invention, the policy function can be selected from deep deterministic policy gradient, proximal policy optimization algorithm; or an optional multi-armed bandit algorithm, where each "arm" is defined as a set of fixed resource allocation strategies, and the current optimal strategy is selected through historical reward evaluation; subsequently, in step S505, according to the resource allocation ratio output by the policy function, dynamic optimization allocation of computing resources is realized.

[0048] This application discloses a hierarchical decision-making system for an unmanned aerial vehicle intelligent agent at the edge and cloud, including: Unmanned aerial vehicle terminal module: Built-in lightweight deep learning model, model compression module, and dynamic task offloading module based on self-attention mechanism, used to process emergency or simple tasks in real time and adaptively offload complex tasks according to task characteristics; Edge computing node module: Used to deploy a lightweight local decision-making model obtained through knowledge distillation, preprocess and make local decisions on the tasks offloaded by the terminal, and determine whether the tasks are further forwarded to the cloud; Cloud data center module: Configured with a large-scale pre-trained model based on the Transformer structure, used to execute global optimization decisions for computationally intensive or data-intensive tasks forwarded by the edge computing node; Dynamic load perception and resource scheduling module: Used to monitor the network and computing load status in real time, and realize dynamic scheduling and optimized allocation of computing resources among the unmanned aerial vehicle terminal, edge computing node, and cloud data center based on a multi-objective optimization algorithm.

[0049] Among them, the dynamic load perception and resource scheduling module adopts a reinforcement learning algorithm, dynamically adjusts the computing load ratio among the terminal, edge, and cloud according to the real-time network status and the load of each node, and can achieve a balance between real-time response and decision-making accuracy.

[0050] In summary, the UAV intelligent agent edge-cloud hierarchical decision-making method and system involved in the embodiments of the present invention can solve the problem that "most of the current technical solutions for UAV tasks using the edge-cloud collaborative computing architecture are difficult to flexibly adapt to network environment changes and computational load fluctuations, lack dynamic decision-making and resource scheduling capabilities, and limit the real-time performance and decision-making accuracy of UAV task execution". For the UAV intelligent agent edge-cloud hierarchical decision-making method disclosed in the present invention, the decision-making method is divided into three layers, specifically the three-layer structure of "UAV terminal - edge computing node - cloud". The tasks are hierarchically processed according to the complexity and real-time requirements of the computing tasks to achieve efficient, real-time, and intelligent task decision-making with limited computing power. Among them, the edge-cloud hierarchical decision-making method mainly includes five major steps: The first step is real-time preliminary decision-making at the terminal, that is, using a compressed lightweight deep learning model on the UAV terminal to perform rapid preliminary decision-making on simple or urgent tasks to meet real-time requirements; the second step is task offloading at the terminal. To meet the real-time processing requirements of complex tasks, some tasks received by the UAV terminal are offloaded to the edge computing node for processing. In this step, a dynamic task offloading mechanism in the UAV terminal is used, and the resource requirements of the tasks are modeled through a self-attention mechanism, and the tasks are offloaded to the edge computing node to achieve task preprocessing and local decision-making; the third step is task processing at the edge computing node, that is, using knowledge distillation technology to obtain a local decision-making model, and the edge computing node uses the local decision-making model to process the offloaded tasks and further determines whether the tasks need to be forwarded to the cloud data center; the fourth step is global optimization decision-making at the cloud. The cloud receives computationally intensive or data-intensive tasks forwarded from the edge computing node. For computationally intensive or data-intensive tasks, a large-scale pre-trained model based on Transformer is used for global optimization decision-making, which can improve the decision-making accuracy of complex task processing; the last step is dynamic load perception and resource scheduling, that is, by real-time monitoring the network bandwidth, latency, and load status of each computing node, dynamically adjusting the allocation ratio of computing resources among the terminal, edge, and cloud, optimizing the allocation of computing resources, and achieving a dynamic balance between the accuracy and real-time response of task decision-making, being able to flexibly adapt to network environment changes and computational load fluctuations, and improving the ability of dynamic decision-making and resource scheduling. It can be seen that the technical solution involved in the present invention, compared with the prior art, can effectively alleviate the computing power and energy consumption bottlenecks of the UAV terminal, dynamically optimize the allocation of computing resources, improve the real-time response ability and decision-making accuracy of UAV complex task processing, and can be widely applied to scenarios such as UAV task planning, target tracking, obstacle avoidance, and inspection and monitoring, with significant practical value.

[0051] The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An edge-cloud hierarchical decision-making method for UAV intelligent agents, characterized in that It includes the following steps: S1. Terminal real-time preliminary decision-making: Construct a lightweight deep learning model, deploy the compressed lightweight deep learning model to the UAV terminal after compression processing, and the terminal makes a quick preliminary decision on simple or urgent tasks based on the compressed lightweight deep learning model; S2. Terminal dynamic task offloading: Build a dynamic task offloading mechanism in the UAV terminal, model the resource requirements of the current task according to task complexity, data volume and network status, and combine the self-attention mechanism, and adaptively offload some computing tasks from the terminal to the edge computing node; S3. Edge computing node task processing: The edge computing node obtains a local decision-making model through knowledge distillation, processes the tasks offloaded from the terminal, and further determines whether the tasks need to be forwarded to the cloud data center; S4. Cloud global optimization decision-making: Forward the tasks determined by the edge computing node as computationally intensive or data-intensive to the cloud data center, and the cloud data center makes global optimization decisions based on the large-scale pre-trained model; S5. Dynamic load awareness and resource scheduling: Real-time monitor and schedule the computing resources of tasks through the dynamic load awareness mechanism among the terminal, edge and cloud layers, optimize the overall resource allocation, and dynamically balance decision-making accuracy and real-time response.

2. The method for hierarchical decision-making of the drone intelligent agent edge-cloud according to claim 1, wherein, In step S1, the method for compressing the lightweight deep learning model includes, but is not limited to, one or more combinations of network pruning, low-rank decomposition, and quantization-aware training.

3. The method for hierarchical decision-making of the drone intelligent agent edge-cloud according to claim 2, wherein In step S1, the steps for compressing the lightweight deep learning model specifically include: S101. Train the original deep learning teacher model in the cloud data center to obtain initial network parameters; S102. Prune redundant parameters of the teacher model using network pruning; the redundant parameter pruning includes: calculating the L1 norm of network weights: , where represents the value of the -th weight in the initial network. Subsequently, set a threshold . Weights with importance lower than the threshold are set to zero to determine the importance of the weights. S103. Perform low-rank decomposition on the model after network pruning, and decompose the model weight matrix through singular value decomposition for processing, and the decomposition formula is: ； Among them, is the left singular matrix retaining the first columns, is the diagonal matrix containing the first singular values, is the right singular matrix retaining the first columns; S104. Perform quantization-aware training on the model compressed by low-rank decomposition, and map the floating-point parameters to the integer space through the quantization function to complete model compression. The implementation formula is as follows: ； Among them, is the quantization scale factor, is the quantization zero point.

4. The method for hierarchical decision-making of the UAV intelligent agent edge-cloud according to claim 1, wherein, In step S2, the method for adaptively offloading some computing tasks from the terminal to the edge computing node is: Generate the attention weights of task features based on the self-attention mechanism, evaluate the urgency and processing priority of task offloading, divide the dynamic computing load between the terminal side and the edge side accordingly, and execute the offloading decision according to the dynamic computing load, and offload some computing tasks from the terminal to the edge computing node. Its specific steps include the following: S201. Task feature extraction and encoding: Extract the feature vector of the task received by the UAV terminal, denoted as the task feature vector , where the task feature vector includes the data volume size, computational complexity, network status, and urgency of the task. Subsequently, task feature embedding is performed on the task feature vector through linear mapping: ； Among them, and are learnable parameter matrices and bias terms; S202. Self-attention mechanism calculates attention weights: Use the scaled dot and attention mechanism in the self-attention mechanism to process the task feature embeddings, calculate the attention weights of the current task features, and obtain the attention vector; ； wherein represents the current task feature and represents the set of historical task features represents the dimension of the feature vector S203. Calculate the urgency and priority of task offloading: Based on the attention vector, evaluate the urgency of the task in combination with a non-linear activation function and the processing priority : ； S204. Dynamic task offloading decision: Based on the urgency level and processing priority determine the strategy and proportion of task offloading. Subsequently, divide the dynamic computing loads of the edge device side and the edge side based on the strategy and proportion of task offloading, and execute the task offloading decision according to the dynamic computing loads; S205. Dynamically adjust the task offloading ratio: The preset task offloading ratio is , and then determine the task offloading ratio based on the urgency : ； When the task urgency level is high, that is then indicating that all tasks are executed locally; When the task urgency level is low, that is , it means that most or all of the task is unloaded.

5. The method for hierarchical decision-making of the drone intelligent agent edge-cloud according to claim 4, characterized in that In step S204, the steps for executing the task offloading decision specifically include: Set threshold , including priority threshold , complexity threshold , data volume threshold ; If the task offloading urgency exceeds the priority threshold , it means the task is urgent and needs to be processed locally immediately without task offloading; If the task priority is low, the complexity is high, or the data volume exceeds the data volume threshold , then unload the task to the edge computing node; The specific decision is: 。 6. The method for hierarchical decision-making of the UAV intelligent agent edge-cloud according to claim 1, wherein In step S3, the method for obtaining the local decision-making model is: Based on knowledge distillation, train a high-precision teacher model in the cloud data center in advance, and then use the edge computing node to guide the training of the lightweight student model with the teacher model, so as to quickly complete task preprocessing and local decision-making on the edge side.

7. The method for hierarchical decision-making of the drone intelligent agent edge-cloud according to claim 6, characterized in that In step S3, the steps for obtaining the local decision-making model specifically include: S301. Train the teacher model: Train a high-precision teacher model in the cloud data center , to minimize the standard classification or regression loss: ； Among them, is the input task data, is the corresponding true label, is the total number of categories, is the prediction probability of the teacher model for the th category; S302. Obtain a soft label: Adjust the temperature of the logit vector output by the teacher model to generate a soft label probability distribution: ； Among them, is the logit output of the teacher model, which is the score before softmax; is the soft label after temperature processing; is the temperature parameter; S303. Train the student model: Deploy a lightweight student model on the edge computing node , and construct a joint distillation loss function according to the soft labels provided by the teacher model as guidance information ； where, is the cross-entropy loss of the student model based on the true labels, is the KL divergence between the student and the teacher, , are the softmax outputs of the student and the teacher respectively after using the same temperature parameter, is the weight coefficient, is the temperature squared term; S304. Model compression and edge deployment: Further perform model compression operations on the student model and deploy it in the edge computing node to quickly preprocess and make local decisions on the tasks offloaded from the terminal.

8. The method for hierarchical decision-making of the UAV intelligent agent edge-cloud according to claim 1, characterized in that In step S4, the large-scale pre-trained model adopts a Transformer structure and is pre-trained using the large-scale environmental data accumulated in the cloud; and the steps for making global optimization decisions based on the large-scale pre-trained model specifically include: S401. Construct the input data format: Structurally process the complex task data uploaded by the drone to construct an input sequence , which is expressed as: ； Among them, represents the input task feature vector, is the positional encoding; S402. Transformer encoding layer processing: The input sequence is processed through a Transformer encoder in multiple layers, each layer including a multi-head self-attention mechanism and a feed-forward network, and the attention calculation: ； Subsequently, multiple heads perform attention calculations in parallel and then splicing processing: ； Then, a two-layer linear transformation and activation function processing are performed on the output vector of each position through a feed-forward network: ； S403. Design of pre-training tasks: The model is pre-trained based on the large-scale environmental data accumulated in the cloud, and the training objectives include but are not limited to path planning loss, image recognition loss, and delay prediction loss. The total loss function is expressed as: ； Among them, , , respectively represent the path optimization loss, the image recognition accuracy loss, and the delay prediction error loss; S404. Decision output in the inference stage: In the inference stage, based on the pre-trained Transformer model, global optimization decisions are made for the computationally intensive or data-intensive tasks forwarded by the edge computing nodes, and the outputs include but are not limited to path prediction results, target recognition categories, or control strategies.

9. The method for hierarchical decision-making of the UAV intelligent agent edge-cloud according to claim 1, wherein In step S5, the method for optimizing the overall resource allocation is: According to the dynamic load perception mechanism, the network bandwidth, delay, and load status of each computing node are monitored in real time, and then a multi-objective optimization algorithm is used to dynamically adjust the computing resource allocation ratio among the terminal, edge, and cloud, specifically including the following steps: S501. Construct the state space: Construct the system state vector , and monitor the network bandwidth in real time , communication delay , and the resource utilization rates of the three types of computing nodes, namely the edge, cloud, and terminal nodes, which is expressed as: ； Among them, , , respectively represent the CPU / GPU resource utilization rates on the terminal, edge, and cloud sides; S502. Define the action space: Define the resource scheduling action space , which represents the computing resource allocation ratios of the three types of nodes, namely the edge, edge-cloud, and cloud nodes, and satisfies the following conditions: ； Among them, , , are the computing resource ratios allocated to the terminal, edge, and cloud respectively; S503. Design the reward function: Using decision delay , energy consumption and task processing accuracy as the optimization objectives, construct a multi-objective optimization reward function : ； Among them, is the weighting coefficient; S504. Reinforcement learning decision-making mechanism: Adopt a resource scheduling strategy based on reinforcement learning to optimize the reward function: ； Among them, is the discount factor, is the policy function; S505. Execute resource scheduling: According to the resource allocation ratio output by the policy function , control the distribution and scheduling of computing tasks among the edge, cloud, and device layers, and dynamically optimize the allocation of computing resources.

10. An intelligent agent edge-cloud hierarchical decision-making system for unmanned aerial vehicles, characterized in that, Including: Drone terminal module: Built-in with a lightweight deep learning model, a model compression module, and a dynamic task offloading module based on the self-attention mechanism, used to process emergency or simple tasks in real time and adaptively offload complex tasks according to task characteristics; Edge computing node module: Used to deploy a lightweight local decision-making model obtained through knowledge distillation, preprocess and make local decisions on the tasks offloaded by the terminal, and determine whether the tasks are further forwarded to the cloud; Cloud data center module: Configured with a large-scale pre-trained model based on the Transformer structure, used to perform global optimization decisions for the computationally intensive or data-intensive tasks forwarded by the edge computing nodes; Dynamic load perception and resource scheduling module: Used to monitor the network and computing load status in real time, and realize dynamic scheduling and optimal allocation of computing resources among the drone terminal, edge computing node, and cloud data center based on a multi-objective optimization algorithm.

Images