A UAV intelligent agent end-edge cloud layered decision-making method and system

Abstract

The present invention discloses a method and system for the end-edge-cloud layered decision-making of an intelligent drone. The decision-making method includes: S1, deploying a compressed lightweight deep learning model on the drone terminal to perform real-time preliminary decisions on simple or urgent tasks; S2, building a dynamic task offloading mechanism to offload complex tasks from the drone terminal to the edge computing node; S3, the edge computing node pre-processing and making local decisions on the task, and further determining whether the task needs to be forwarded to the cloud data center; S4, the cloud data center making global optimization decisions for computationally intensive or data-intensive tasks; S5, real-time monitoring of network and computing node status, and dynamic adjustment of end, edge, and cloud computing resource allocation to achieve a dynamic balance between decision-making accuracy and real-time response. The technical solution disclosed in the present invention can effectively alleviate the computing power and energy consumption bottlenecks of the drone terminal, dynamically optimize computing resource allocation, and improve the real-time response capability and decision-making accuracy of the drone in processing complex tasks.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and unmanned aerial vehicle (UAV) technology, and more specifically, to a UAV intelligent body end-edge-cloud layered decision-making method and system. Background Art

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

[0003] At present, in order to meet the real-time processing requirements of complex drone tasks, the existing technology has proposed an end-edge-cloud collaborative computing architecture. The end-edge-cloud collaborative computing architecture is a distributed computing framework that organically combines terminal devices, edge computing nodes, and cloud computing centers. It aims to meet the diverse scenario requirements of low latency, high reliability, and large bandwidth through resource collaboration, task scheduling, and data processing optimization. When a drone performs a complex task, efficient processing is achieved by rationally allocating the task to the drone terminal, edge computing nodes, and cloud data centers. However, the current technical solutions using the end-edge-cloud collaborative computing architecture are often difficult to flexibly adapt to changes in the network environment and fluctuations in computing load, and lack the ability to make dynamic decisions and resource scheduling, which limits the real-time performance and decision-making accuracy of drone task execution.

[0004] Therefore, how to provide a drone intelligent end-edge-cloud layered decision-making method and system that can effectively alleviate the computing power and energy consumption bottlenecks of drone terminals, dynamically optimize computing resource allocation, and improve the real-time response capability and decision-making accuracy of drone complex task processing has become a technical problem that needs to be urgently solved by technical personnel in this field. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides a drone intelligent body end-edge-cloud hierarchical decision-making method and system, which can effectively alleviate the computing power and energy consumption bottlenecks of drone terminals, dynamically optimize computing resource allocation, and improve the real-time response capability and decision-making accuracy of drone complex task processing.

[0006] The technical solutions provided by the present invention are as follows:

[0007] The present invention provides a UAV intelligent agent terminal-edge-cloud layered decision method, comprising the following steps:

[0008] S1. Real-time preliminary decision-making by the terminal: A lightweight deep learning model is constructed, compressed, and deployed to the UAV terminal. The terminal makes quick preliminary decisions on simple or urgent tasks based on the compressed lightweight deep learning model.

[0009] S2. Dynamic task offloading at the terminal: A dynamic task offloading mechanism is built in the UAV terminal. Based on the task complexity, data volume, and network status, the resource requirements of the current task are modeled in conjunction with the self-attention mechanism, and some computing tasks are adaptively offloaded from the terminal to the edge computing node.

[0010] The method for adaptively offloading part of the computing tasks from the terminal to the edge computing node is:

[0011] The self-attention mechanism generates attention weights for task features, evaluates the urgency and processing priority of task offloading, and divides the dynamic computing load between the terminal and edge. Based on the dynamic computing load, an offloading decision is made to offload some computing tasks from the terminal to the edge computing node. The specific steps include:

[0012] S201, task feature extraction and encoding: Extract the feature vector of the task received by the UAV terminal and record it as the task feature vector , the task feature vector includes the data size, computational complexity, network status and urgency of the task, and then the task feature vector is embedded into the task feature through linear mapping:

[0013] ;

[0014] in, and is the learnable parameter matrix and bias term;

[0015] S202. Calculate the attention weight using the self-attention mechanism: Use the scaling point and attention mechanism in the self-attention mechanism to process the task feature embedding, calculate the attention weight of the current task feature, and obtain the attention vector:

[0016] ;

[0017] in Indicates the current task characteristics, and represents the historical task feature set, represents the feature vector dimension;

[0018] S203, calculate the urgency and priority of task offloading: evaluate the urgency of the task based on the attention vector and combined with the nonlinear activation function and processing priority :

[0019] ;

[0020] S204, dynamic task offloading decision: according to the urgency and processing priority Determining a task offloading strategy and ratio, then dividing the dynamic computing load on the end side and the edge side based on the task offloading strategy and ratio, and executing a task offloading decision based on the dynamic computing load;

[0021] In step S204, the steps of executing task offloading decision specifically include: setting a threshold , including priority thresholds , complexity threshold , data volume threshold ; If the task is unloaded urgently Exceeding the priority threshold , indicating 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 the task is offloaded to the edge computing node; the specific decision is:

[0022] ;

[0023] S205, dynamic adjustment of task offloading ratio: the preset task offloading ratio is , then based on the urgency Determine the task offloading ratio :

[0024] ;

[0025] When the task urgency is high, When , indicating that all tasks are executed locally;

[0026] When the task urgency is low, , indicating that most or all tasks are offloaded;

[0027] S3, Edge computing node task processing: 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;

[0028] S4. Cloud-based global optimization decision: Tasks determined by edge computing nodes to be compute-intensive or data-intensive are forwarded to the cloud data center, which then makes a global optimization decision based on a large-scale pre-trained model. The large-scale pre-trained model uses a Transformer structure and is pre-trained using large-scale environmental data accumulated in the cloud.

[0029] S5. Dynamic load perception and resource scheduling: Through the dynamic load perception mechanism between the end, edge, and cloud layers, the computing resources of the tasks are monitored and scheduled in real time, the overall resource allocation is optimized, and the decision accuracy and real-time response are dynamically balanced.

[0030] Furthermore, in a preferred embodiment of the present invention, in step S1, the method of 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.

[0031] Furthermore, in a preferred embodiment of the present invention, in step S1, the step of compressing the lightweight deep learning model specifically includes:

[0032] S101. Train the original deep learning teacher model in the cloud data center to obtain initial network parameters;

[0033] S102, using network pruning to prune redundant parameters of the teacher model; redundant parameter pruning includes calculating the L1 norm of the network weight: ,in Indicates the initial network The value of the weight, and then set the threshold , the importance is lower than the threshold The weights are reset to zero to determine the importance of the weights;

[0034] S103, perform low-rank decomposition on the model after network pruning, and perform singular value decomposition on the model weight matrix Decomposition is performed, and the decomposition formula is:

[0035] ;

[0036] in, Before retaining The left singular matrix of columns, Is included before A diagonal matrix of singular values, Before retaining Right singular matrix of columns;

[0037] S104, implement quantization-aware training on the model after low-rank decomposition and compression, through the quantization function Mapping floating-point parameters to integer space to complete model compression is achieved with the following formula:

[0038] ;

[0039] in, is the quantization scale factor, is the quantized zero point.

[0040] 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, a high-precision teacher model is pre-trained in the cloud data center, and then the 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.

[0041] Furthermore, in a preferred embodiment of the present invention, in step S3, the step of obtaining the local decision model specifically includes:

[0042] S301, training teacher model: training a high-precision teacher model in the cloud data center , to minimize the standard classification or regression loss:

[0043] ;

[0044] in, To input task data, is the corresponding true label, is the total number of categories, For the teacher model The predicted probability of the class;

[0045] S302, get soft label: the logit vector output by the teacher model Perform temperature adjustment to generate soft label probability distribution:

[0046] ;

[0047] in, is the logit output of the teacher model, which is the score before softmax; It is a soft label after temperature treatment; is the temperature parameter;

[0048] S303. Training student model: deploying lightweight student model on edge computing nodes , based on the soft labels provided by the teacher model as guidance information, a joint distillation loss function is constructed:

[0049] ;

[0050] in, 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 after using the same temperature parameters for students and teachers, Weight coefficient, is the square term of temperature;

[0051] 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 tasks offloaded from the terminal.

[0052] Furthermore, in a preferred embodiment of the present invention, in step S4, the step of performing global optimization decision-making based on a large-scale pre-trained model specifically includes:

[0053] S401, construct input data format: structure the complex mission data uploaded by the drone and construct the input sequence , expressed as:

[0054] ;

[0055] in, represents the input task feature vector, Encode for position;

[0056] S402, Transformer encoding layer processing: The input sequence is processed in multiple layers through the Transformer encoder. Each layer includes a multi-head self-attention mechanism and a feedforward network. Attention calculation:

[0057] ;

[0058] Then multiple heads perform attention calculations in parallel and then perform splicing processing:

[0059] ;

[0060] The output vector of each position is then processed by two layers of linear transformation and activation function through the feedforward network:

[0061] ;

[0062] S403. Design pre-training tasks: Pre-train the model based on 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:

[0063] ;

[0064] in, , , They represent path optimization loss, image recognition accuracy loss, and delay prediction error loss respectively;

[0065] S404, decision output in the reasoning phase: In the reasoning phase, a pre-trained Transformer model is used to make global optimization decisions on the computationally intensive or data-intensive tasks forwarded by the edge computing nodes. The output includes but is not limited to path prediction results, target recognition categories, or control strategies.

[0066] Furthermore, in a preferred embodiment of the present invention, in step S5, the method for optimizing overall resource allocation is: monitoring network bandwidth, latency, and the load status of each computing node in real time based on a dynamic load perception mechanism, and then using a multi-objective optimization algorithm to dynamically adjust the computing resource allocation ratio between the end, edge, and cloud. Specifically, the method includes the following steps:

[0067] S501. Construct state space: Construct system state vector , real-time monitoring of network bandwidth , communication delay , and the resource utilization of the three types of computing nodes, namely, end, edge, and cloud, are expressed as:

[0068] ;

[0069] in, , , Represents the CPU / GPU resource utilization on the device, edge, and cloud sides respectively;

[0070] S502. Define action space: define resource scheduling action space , which represents the computing resource allocation ratio of the three types of nodes: end, edge, and cloud, so that it satisfies:

[0071] ;

[0072] in, , , The proportion of computing resources allocated to terminals, edges, and clouds respectively;

[0073] S503. Design reward function: decision delay , energy consumption and task processing accuracy To optimize the goal, construct a multi-objective optimization reward function :

[0074] ;

[0075] in, is the weighting coefficient;

[0076] S504, reinforcement learning decision-making mechanism: adopting a resource scheduling strategy based on reinforcement learning to optimize the reward function:

[0077] ;

[0078] in, is the discount factor, is the policy function;

[0079] S505, perform resource scheduling: allocate resources according to the output of the policy function , control the distribution and scheduling of computing tasks among the end, edge, and cloud layers, and dynamically optimize the allocation of computing resources.

[0080] In addition, this application also provides another technical solution:

[0081] The present application discloses a drone intelligent end-edge-cloud hierarchical decision-making system, which includes: a drone terminal module: a built-in lightweight deep learning model, a model compression module, and a dynamic task offloading module based on a self-attention mechanism, which is used to process urgent 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 model obtained based on knowledge distillation, pre-process and make local decisions on terminal offloaded tasks, 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 a Transformer structure, which is used to execute global optimization decisions for computationally intensive or data-intensive tasks forwarded by edge computing nodes; 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 optimal allocation of computing resources between drone terminals, edge computing nodes and cloud data centers based on a multi-objective optimization algorithm.

[0082] The present invention provides a method and system for the layered decision-making of an intelligent end-edge-cloud by a drone, wherein the method comprises the following steps: S1, real-time preliminary decision-making by the terminal: constructing a lightweight deep learning model, compressing the lightweight deep learning model and deploying it to the drone terminal, and the terminal performs a quick preliminary decision on simple or urgent tasks based on the compressed lightweight deep learning model; S2, dynamic task offloading by the terminal: constructing a dynamic task offloading mechanism in the drone terminal, modeling the resource requirements of the current task according to the complexity of the task, the amount of data and the network status, and combining the self-attention mechanism, and adaptively offloading part of the computing tasks from the terminal to the edge. Edge computing nodes; S3, edge computing node task processing: edge computing nodes obtain local decision models through knowledge distillation, process tasks unloaded from the terminal, and further determine whether the tasks need to be forwarded to the cloud data center; S4, cloud global optimization decision: tasks judged by edge computing nodes as compute-intensive or data-intensive are forwarded to the cloud data center, and the cloud data center makes global optimization decisions based on large-scale pre-trained models; S5, dynamic load perception and resource scheduling: through the dynamic load perception mechanism between the end, edge, and cloud layers, the computing resources of the tasks are monitored and scheduled in real time, the overall resource allocation is optimized, and the decision accuracy and real-time response are dynamically balanced.Regarding the UAV intelligent body end-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", and the tasks are hierarchically processed according to the complexity and real-time requirements of the computing tasks, so as to achieve efficient, real-time and intelligent task decision-making under limited computing power; wherein, the end-edge-cloud hierarchical decision-making method mainly includes five steps: the first step is the real-time preliminary decision of the terminal, that is, using a compressed lightweight deep learning model on the UAV terminal, and using the lightweight deep learning model to perform quick preliminary decisions on simple or urgent tasks to meet the real-time requirements; the second step is terminal task offloading. In order to meet the real-time processing requirements of complex tasks, part of the tasks received by the UAV terminal are offloaded to the edge computing node for processing. In this step, the dynamic task offloading mechanism in the UAV terminal under construction is utilized, and the resource requirements required for the task are modeled through the self-attention mechanism, and the task is offloaded to the edge computing node to achieve Task preprocessing and local decision-making; the third step is edge computing node task processing, that is, using knowledge distillation technology to obtain a local decision model, and the edge computing node uses the local decision model to process the offloaded task, and further determines whether the task needs to be forwarded to the cloud data center; the fourth step is cloud-based global optimization decision-making, the cloud receives computationally intensive or data-intensive tasks forwarded from the edge computing node, and 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 of network bandwidth, delay and the load status of each computing node, dynamically adjusting the computing resource allocation ratio between the end, edge and cloud, optimizing computing resource allocation, and achieving a dynamic balance between the accuracy of task decision-making and real-time response, which can flexibly adapt to changes in the network environment and fluctuations in computing load, and improve the ability of dynamic decision-making and resource scheduling. It can be seen that compared with the existing technology, the technical solution involved in the present invention can effectively alleviate the computing power and energy consumption bottlenecks of drone terminals, dynamically optimize the allocation of computing resources, and improve the real-time response capability and decision-making accuracy of drone complex task processing. It can be widely used in scenarios such as drone mission planning, target tracking, obstacle avoidance, and inspection and monitoring, and has significant practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0084] Figure 1This is a flowchart of the steps of the UAV intelligent agent end-edge cloud layered decision method involved in an embodiment of the present invention;

[0085] Figure 2 A flowchart of the steps of a method for compressing the lightweight deep learning model according to an embodiment of the present invention;

[0086] Figure 3 This is a flowchart of the steps of adaptively offloading part of the computing tasks from the terminal to the edge computing node involved in an embodiment of the present invention;

[0087] Figure 4 A flowchart of the steps of obtaining a local decision model according to an embodiment of the present invention;

[0088] Figure 5 This is a flowchart of the steps for performing global optimization decision-making based on a large-scale pre-trained model according to an embodiment of the present invention;

[0089] Figure 6 This is a flow chart of the steps for optimizing overall resource allocation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0090] In order to enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0091] It should be noted that when an element is referred to as being “fixed on” or “set on” another element, it can be directly on the other element or indirectly set on the other element; when an element is referred to as being “connected to” another element, it can be directly connected to the other element or indirectly connected to the other element.

[0092] It should be understood that the terms "length", "width", "up", "down", "front", "back", "first", "second", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as limiting the present invention.

[0093] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "plurality" or "several" means two or more, unless otherwise specifically defined.

[0094] It should be noted that the structures, proportions, sizes, etc. depicted in the drawings of this specification are only used to match the contents disclosed in the specification so that people familiar with this technology can understand and read them. They are not used to limit the conditions under which the present invention can be implemented. Therefore, they have no substantive technical significance. Any structural modification, change in proportional relationship or adjustment of size should still fall within the scope of the technical content disclosed in the present invention without affecting the efficacy and purpose that can be achieved by the present invention.

[0095] Please Figures 1 to 6As shown, the present invention provides a UAV intelligent end-edge-cloud hierarchical decision-making method and system, which integrates model compression, self-attention mechanism, knowledge distillation, Transformer pre-training model and dynamic load perception mechanism, and can optimize the computing resource scheduling between UAV terminals, edge computing nodes and cloud data centers. In the UAV intelligent end-edge-cloud hierarchical decision-making method, the UAV's intelligent decision-making is divided into a "three-layer structure of end, edge and cloud", and the computing tasks are hierarchically processed according to the task complexity and real-time requirements, which can achieve efficient, real-time and intelligent task decision-making under limited computing power. It specifically includes the following steps: S1, terminal real-time preliminary decision: construct a lightweight deep learning model, compress the lightweight deep learning model and deploy it to the UAV terminal, and the terminal performs quick preliminary decision 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 UAV terminal, according to the task complexity, data volume and network status, and combines the self-attention mechanism to model the resource requirements of the current task, 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 model through knowledge distillation, processes the task unloaded from the terminal, and further determines whether the task needs to be forwarded to the cloud data center; S4, cloud global optimization decision: forward the tasks judged 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 perception and resource scheduling: through the dynamic load perception mechanism between the end, edge, and cloud layers, the computing resources of the task are monitored and scheduled in real time, the overall resource allocation is optimized, and the decision accuracy and real-time response are dynamically balanced. 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 allocation of computing resources, and improve the real-time response capability and decision accuracy of the drone's complex task processing.

[0096] The following is a detailed description of the drone intelligent agent end-edge-cloud layered decision-making method disclosed in the present invention with reference to specific embodiments. The drone intelligent agent end-edge-cloud layered decision-making method specifically includes the following steps:

[0097] S1. Real-time preliminary decision-making by the terminal: A lightweight deep learning model is constructed, compressed, and deployed to the UAV terminal. The terminal performs quick preliminary decisions on simple or urgent tasks based on the compressed lightweight deep learning model.

[0098] Among them, in a specific embodiment of the present invention, step S1 is used to realize real-time decision-making task processing of the terminal device under the constraints of computing power and energy consumption; in this step, a lightweight deep learning model is used on the drone terminal, and then the model compression technology is used to compress the lightweight deep learning model and deploy it on the drone terminal, so as to realize rapid preliminary decision-making for simple or complex tasks by using the lightweight deep learning model to meet real-time requirements.

[0099] Specifically, in a specific embodiment of the present invention, in step S1, the method of 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.

[0100] Specifically, in step S1, after the lightweight deep learning model is constructed on the UAV terminal, the model needs to be compressed. The model compression processing aims to optimize the model structure and parameters so that it can significantly reduce resource consumption and improve efficiency while maintaining high performance. After compression processing, 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 an embodiment of the present invention, the method of compressing the lightweight deep learning model includes one or more combinations of network pruning, low-rank decomposition and quantization-aware training.

[0101] Specifically, in a specific embodiment of the present invention, in step S1, the step of compressing the lightweight deep learning model specifically includes: S101, training the original deep learning teacher model in a cloud data center to obtain initial network parameters; S102, using network pruning to prune redundant parameters of the teacher model; redundant parameter pruning includes calculating the L1 norm of the network weight:

[0102] ;

[0103] in, Indicates the initial network The value of the weight, and then set the threshold , the importance is lower than the threshold The weights are reset to zero to determine the importance of the weights; S103, low-rank decomposition is performed on the model after network pruning, and the model weight matrix is decomposed by singular value decomposition. Decomposition is performed, and the decomposition formula is:

[0104] ;

[0105] in, Before retaining The left singular matrix of columns, Is included before A diagonal matrix of singular values, Before retaining S104, implement quantization-aware training on the model after low-rank decomposition and compression, through the quantization function Mapping floating-point parameters to integer space to complete model compression is achieved with the following formula:

[0106] ;

[0107] in, is the quantization scale factor, is the quantized zero point.

[0108] In an embodiment of the present invention, step S1 may perform model compression using one or more combinations of network pruning, low-rank decomposition, and quantization-aware training techniques; 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, eliminate 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 retaining The left singular matrix of columns, Is included before A diagonal matrix of singular values, Before retaining The right singular matrix of the column; keep the first The largest singular value, and retain the matrix accordingly 、 、 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 the model parameters, and utilizes adaptive optimization in the training phase to ensure 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 UAV terminal to realize real-time decision-making task processing of the terminal device under the constraints of computing power and energy consumption.

[0109] S2. Dynamic task offloading at the terminal: A dynamic task offloading mechanism is built in the drone terminal. Based on 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.

[0110] Among them, in an embodiment of the present invention, step S2 is terminal task offloading. In order to meet the real-time processing requirements of complex tasks, part of the tasks received by the drone terminal are offloaded to the edge computing node for processing. In this step, the dynamic task offloading mechanism in the drone terminal under construction is utilized, and the resource requirements required for the task are modeled through the self-attention mechanism, and the task is offloaded to the edge computing node, so that it can realize task preprocessing and local decision-making.

[0111] Specifically, in a specific 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: generating the attention weight of the task feature based on the self-attention mechanism, evaluating the urgency and processing priority of the task offloading, thereby dividing the dynamic computing load on the terminal side and the edge side, executing the offloading decision according to the dynamic computing load, and offloading part of the computing tasks from the terminal to the edge computing node.

[0112] Among them, in an embodiment of the present invention, the offloading of tasks from the drone terminal to the edge computing node is based on dynamic computing load division. The self-attention mechanism in step S2 generates attention weights of task features to evaluate the urgency and processing priority of task offloading, thereby realizing dynamic computing load division between the terminal side and the edge side.

[0113] Specifically, in a specific embodiment of the present invention, the step of adaptively offloading part of the computing tasks from the terminal to the edge computing node specifically includes: S201, task feature extraction and encoding: extracting the feature vector of the task received by the drone terminal, recorded as the task feature vector , the task feature vector includes the data size, computational complexity, network status and urgency of the task, and then the task feature vector is embedded into the task feature through linear mapping:

[0114] ;

[0115] in, and is the learnable parameter matrix and bias term; S202, self-attention mechanism calculates attention weight: the scaling point and attention mechanism in the self-attention mechanism are used to process the task feature embedding, calculate the attention weight of the current task feature, and obtain the attention vector:

[0116] ;

[0117] in, Indicates the current task characteristics, and represents the historical task feature set, Represents the feature vector dimension; S203, calculate the urgency and priority of task offloading: evaluate the urgency of the task based on the attention vector and combined with the nonlinear activation function and processing priority :

[0118] ;

[0119] S204, dynamic task offloading decision: according to the urgency and processing priority Determining a task offloading strategy and ratio, then dividing the dynamic computing load on the end side and the edge side based on the task offloading strategy and ratio, and executing a task offloading decision based on the dynamic computing load;

[0120] S205, dynamic adjustment of task offloading ratio: the preset task offloading ratio is , then based on the urgency Determine the task offloading ratio :

[0121] ;

[0122] When the task urgency is high, When , indicating that all tasks are executed locally; when the task urgency is low, that is, , indicating that most or all tasks are offloaded.

[0123] Specifically, in a specific embodiment of the present invention, in step S204, the step of executing task offloading decision specifically includes: setting a threshold , including priority thresholds , complexity threshold , data volume threshold ,; If the task offloading urgency Exceeding the priority threshold , indicating 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;

[0124] The specific decisions are:

[0125] .

[0126] S3. Edge computing node task processing: The edge computing node obtains the 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.

[0127] Among them, in an embodiment of the present invention, step S3 edge computing node task processing, that is, using knowledge distillation technology to obtain a local decision model, and the edge computing node uses the local decision model to process the offloaded task and further determine whether the task needs to be forwarded to the cloud data center.

[0128] Specifically, in a specific embodiment of the present invention, in step S3, the method for obtaining a local decision model is: based on knowledge distillation, a high-precision teacher model is pre-trained in a cloud data center, and then the 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.

[0129] Specifically, in a specific embodiment of the present invention, in step S3, the step of obtaining the local decision model specifically includes: S301, training the teacher model: training a high-precision teacher model in the cloud data center , to minimize the standard classification or regression loss:

[0130] ;

[0131] in, To input task data, is the corresponding true label, is the total number of categories, For the teacher model The predicted probability of the class; S302, get the soft label: the logit vector output by the teacher model Perform temperature adjustment to generate soft label probability distribution:

[0132] ;

[0133] in, is the logit output of the teacher model, which is the score before softmax; is a soft label after temperature processing, and the soft label is used to guide the student model learning; is a temperature parameter, which can be used to smooth the probability distribution; S303, training student model: deploying a lightweight student model on the edge computing node , based on the soft labels provided by the teacher model as guidance information, a joint distillation loss function is constructed:

[0134] ;

[0135] in, 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 after using the same temperature parameters for students and teachers, Weight coefficient, which is used to weigh the two parts of loss, It is the temperature square term. Setting the temperature square term can perform gradient adjustment in back propagation; S304, model compression and edge deployment: further perform model compression operation on the student model and deploy it in the edge computing node. The compressed student model is used for fast preprocessing and local decision-making of tasks offloaded from the terminal.

[0136] S4. Global optimization decision-making on the cloud: Tasks judged as compute-intensive or data-intensive by edge computing nodes are forwarded to the cloud data center, which makes global optimization decisions based on a large-scale pre-trained model; the large-scale pre-trained model adopts a Transformer structure and is pre-trained using large-scale environmental data accumulated in the cloud.

[0137] Specifically, in a specific embodiment of the present invention, in step S4, the large-scale pre-training model adopts a Transformer structure and is pre-trained using large-scale environmental data accumulated in the cloud.

[0138] In a specific embodiment of the present invention, in order to improve the decision-making accuracy of complex task processing, in step S4, the cloud data center uses a large-scale pre-trained model based on Transformer to perform global optimization decision-making for computationally intensive or data-intensive tasks.

[0139] Specifically, in a specific embodiment of the present invention, the steps of making global optimization decisions based on a large-scale pre-training model specifically include: S401, constructing an input data format: structuring the complex mission data uploaded by the drone to construct an input sequence , expressed as:

[0140] ;

[0141] in, represents the input task feature vector, Encode for position;

[0142] S402, Transformer encoding layer processing: The input sequence is processed in multiple layers through the Transformer encoder. Each layer includes a multi-head self-attention mechanism and a feedforward network. Attention calculation:

[0143] ;

[0144] Then multiple heads perform attention calculations in parallel and then perform splicing processing:

[0145] ;

[0146] The output vector of each position is then processed by two layers of linear transformation and activation function through the feedforward network:

[0147] ;

[0148] S403. Design pre-training tasks: Pre-train the model based on 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:

[0149] ;

[0150] in, , , They represent path optimization loss, image recognition accuracy loss, and delay prediction error loss respectively; S404, decision output in the reasoning phase: In the reasoning phase, a global optimization decision is made on the computationally intensive or data-intensive tasks forwarded by the edge computing node based on the pre-trained Transformer model, and the output includes but is not limited to path prediction results, target recognition categories, or control strategies.

[0151] S5. Dynamic load perception and resource scheduling: Through the dynamic load perception mechanism between the end, edge, and cloud layers, the computing resources of the tasks are monitored and scheduled in real time, the overall resource allocation is optimized, and the decision accuracy and real-time response are dynamically balanced.

[0152] Specifically, in a specific embodiment of the present invention, in step S5, the method for optimizing the overall resource allocation is: real-time monitoring of network bandwidth, latency and the load status of each computing node based on a dynamic load perception mechanism, and then using a multi-objective optimization algorithm to dynamically adjust the computing resource allocation ratio between the end, edge and cloud.

[0153] Wherein, the multi-objective optimization algorithm in step S5 includes reinforcement learning or a multi-armed bandit algorithm.

[0154] Specifically, in a specific embodiment of the present invention, the step of optimizing the overall resource allocation specifically includes: S501, constructing a state space: constructing a system state vector , real-time monitoring of network bandwidth , communication delay , and the resource utilization of the three types of computing nodes, namely, end, edge, and cloud, are expressed as:

[0155] ;

[0156] in, , , Represents the CPU / GPU resource utilization of the end, edge, and cloud sides respectively; S502, define action space: define resource scheduling action space , which represents the computing resource allocation ratio of the three types of nodes: end, edge, and cloud, so that it satisfies:

[0157] ;

[0158] in, , , The proportion of computing resources allocated to the terminal, edge, and cloud respectively; S503, design a reward function, the reward function is used for reinforcement learning: , energy consumption and task processing accuracy To optimize the goal, construct a multi-objective optimization reward function :

[0159] ;

[0160] in, is a weighted coefficient; S504, reinforcement learning decision mechanism: adopting a resource scheduling strategy based on reinforcement learning to optimize the reward function:

[0161] ;

[0162] in, is the discount factor, is a policy function; S505, executing resource scheduling: allocating resources according to the output of the policy function , control the distribution and scheduling of computing tasks among the end, edge, and cloud layers, and dynamically optimize the allocation of computing resources.

[0163] Specifically, step S504 is used to optimize the reward function through the policy function; in an embodiment of the present invention, the policy function can use a deep deterministic policy gradient or a proximal policy optimization algorithm; or a multi-armed bandit algorithm can be used to define each "arm" as a set of fixed resource allocation strategies, and select the current optimal strategy through historical reward evaluation; then in step S505, it realizes the dynamic optimization allocation of computing resources according to the resource allocation ratio output by the policy function.

[0164] The present application discloses a drone intelligent body end-edge-cloud hierarchical decision-making system, including: a drone terminal module: with a built-in lightweight deep learning model, a model compression module and a dynamic task offloading module based on the self-attention mechanism, which is used to process urgent 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 model obtained based on knowledge distillation, pre-process and make local decisions on terminal offloaded tasks, 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, which is used to execute global optimization decisions for computationally intensive or data-intensive tasks forwarded by edge computing nodes; 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 optimal allocation of computing resources between drone terminals, edge computing nodes and cloud data centers based on a multi-objective optimization algorithm.

[0165] Among them, the dynamic load perception and resource scheduling module adopts a reinforcement learning algorithm to dynamically adjust the computing load ratio between the end, edge, and cloud according to the real-time status of the network and the load of each node, which can achieve a balance between real-time response and decision-making accuracy.

[0166] To sum up, the drone intelligent body end-edge-cloud hierarchical decision-making method and system involved in the embodiments of the present invention can solve the problem that "the current technical solutions for drone missions using end-edge-cloud collaborative computing architecture are mostly difficult to flexibly adapt to changes in network environment and fluctuations in computing load, lack dynamic decision-making and resource scheduling capabilities, and limit the real-time performance and decision-making accuracy of drone mission execution." Regarding the UAV intelligent body end-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", and the tasks are hierarchically processed according to the complexity and real-time requirements of the computing tasks, so as to achieve efficient, real-time and intelligent task decision-making under limited computing power; wherein, the end-edge-cloud hierarchical decision-making method mainly includes five steps: the first step is the real-time preliminary decision of the terminal, that is, using a compressed lightweight deep learning model on the UAV terminal, and using the lightweight deep learning model to perform quick preliminary decisions on simple or urgent tasks to meet the real-time requirements; the second step is terminal task offloading. In order to meet the real-time processing requirements of complex tasks, part of the tasks received by the UAV terminal are offloaded to the edge computing node for processing. In this step, the dynamic task offloading mechanism in the UAV terminal under construction is utilized, and the resource requirements required for the task are modeled through the self-attention mechanism, and the task is offloaded to the edge computing node to achieve Task preprocessing and local decision-making; the third step is edge computing node task processing, that is, using knowledge distillation technology to obtain a local decision model, and the edge computing node uses the local decision model to process the offloaded task, and further determines whether the task needs to be forwarded to the cloud data center; the fourth step is cloud-based global optimization decision-making, the cloud receives computationally intensive or data-intensive tasks forwarded from the edge computing node, and 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 of network bandwidth, delay and the load status of each computing node, dynamically adjusting the computing resource allocation ratio between the end, edge and cloud, optimizing computing resource allocation, and achieving a dynamic balance between the accuracy of task decision-making and real-time response, which can flexibly adapt to changes in the network environment and fluctuations in computing load, and improve the ability of dynamic decision-making and resource scheduling. It can be seen that compared with the existing technology, the technical solution involved in the present invention can effectively alleviate the computing power and energy consumption bottlenecks of drone terminals, dynamically optimize the allocation of computing resources, and improve the real-time response capability and decision-making accuracy of drone complex task processing. It can be widely used in scenarios such as drone mission planning, target tracking, obstacle avoidance, and inspection and monitoring, and has significant practical value.

[0167] The above description of the disclosed embodiments is intended to enable one 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. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A UAV intelligent agent end-edge cloud layered decision-making method, characterized by: The following steps are involved: S1. Real-time preliminary decision-making by the terminal: A lightweight deep learning model is constructed, compressed, and deployed to the UAV terminal. The terminal makes quick preliminary decisions on simple or urgent tasks based on the compressed lightweight deep learning model. S2. Dynamic task offloading at the terminal: A dynamic task offloading mechanism is built in the UAV terminal. Based on the task complexity, data volume, and network status, the resource requirements of the current task are modeled in conjunction with the self-attention mechanism, and some computing tasks are adaptively offloaded from the terminal to the edge computing node. The method for adaptively offloading some computing tasks from the terminal to the edge computing node is as follows: generating attention weights of task features based on a self-attention mechanism, evaluating the urgency and processing priority of task offloading, thereby dividing the dynamic computing load on the terminal side and the edge side, executing offloading decisions based on the dynamic computing load, and offloading some computing tasks from the terminal to the edge computing node. The method specifically includes the following steps: S201, task feature extraction and encoding: Extract the feature vector of the task received by the UAV terminal and record it as the task feature vector , the task feature vector includes the data size, computational complexity, network status and urgency of the task, and then the task feature vector is embedded into the task feature through linear mapping: ； in, and is the learnable parameter matrix and bias term; S202. Calculate the attention weight using the self-attention mechanism: Use the scaling point and attention mechanism in the self-attention mechanism to process the task feature embedding, calculate the attention weight of the current task feature, and obtain the attention vector: ； in Indicates the current task characteristics, and represents the historical task feature set, represents the feature vector dimension; S203, calculate the urgency and priority of task offloading: evaluate the urgency of the task based on the attention vector and combined with the nonlinear activation function and processing priority : ； S204, dynamic task offloading decision: according to the urgency and processing priority Determining a task offloading strategy and ratio, then dividing the dynamic computing load on the end side and the edge side based on the task offloading strategy and ratio, and executing a task offloading decision based on the dynamic computing load; In step S204, the steps of executing task offloading decision specifically include: setting a threshold , including priority thresholds , complexity threshold , data volume threshold ; If the task is unloaded urgently Exceeding the priority threshold , indicating 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 the task is offloaded to the edge computing node; the specific decision is: ； S205, dynamic adjustment of task offloading ratio: the preset task offloading ratio is , then based on the urgency Determine the task offloading ratio : ； When the task urgency is high, When , indicating that all tasks are executed locally; When the task urgency is low, , indicating that most or all tasks are offloaded; S3, Edge computing node task processing: 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; S4, Cloud-based global optimization decision: Forward tasks determined by edge computing nodes to be compute-intensive or data-intensive to the cloud data center, which then makes global optimization decisions based on a large-scale pre-trained model. The large-scale pre-training model adopts the Transformer structure and is pre-trained using large-scale environmental data accumulated in the cloud; S5. Dynamic load perception and resource scheduling: Through the dynamic load perception mechanism between the end, edge, and cloud layers, the computing resources of the tasks are monitored and scheduled in real time, the overall resource allocation is optimized, and the decision accuracy and real-time response are dynamically balanced.

2. The UAV agent-edge-cloud layered decision-making method according to claim 1 is characterized in that: In step S1, the method of 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 UAV agent-edge-cloud layered decision-making method according to claim 2 is characterized in that: 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, using network pruning to prune redundant parameters of the teacher model; the redundant parameter pruning includes: calculating the L1 norm of the network weight: ,in Indicates the initial network The value of the weight, and then set the threshold , the importance is lower than the threshold The weights are reset to zero to determine the importance of the weights; S103, perform low-rank decomposition on the model after network pruning, and perform singular value decomposition on the model weight matrix Decomposition is performed, and the decomposition formula is: ； in, Before retaining The left singular matrix of columns, Is included before A diagonal matrix of singular values, Before retaining Right singular matrix of columns; S104, implement quantization-aware training on the model after low-rank decomposition and compression, through the quantization function Mapping floating-point parameters to integer space to complete model compression is achieved with the following formula: ； in, is the quantization scale factor, is the quantized zero point.

4. The UAV agent-edge-cloud layered decision-making method according to claim 1 is characterized in that: In step S3, the method for obtaining the local decision model is: based on knowledge distillation, a high-precision teacher model is pre-trained in the cloud data center, and then the 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.

5. The UAV agent-edge-cloud layered decision-making method according to claim 4 is characterized in that: In step S3, the steps of obtaining the local decision model specifically include: S301, training teacher model: training a high-precision teacher model in the cloud data center , to minimize the standard classification or regression loss: ； in, To input task data, is the corresponding true label, is the total number of categories, For the teacher model The predicted probability of the class; S302, obtain soft labels: output by the teacher model vector Perform temperature adjustment to generate soft label probability distribution: ； in, For teacher models Output, which is the score before softmax; It is a soft label after temperature treatment; is the temperature parameter; S303. Training student model: deploying lightweight student model on edge computing nodes , based on the soft labels provided by the teacher model as guidance information, a joint distillation loss function is constructed: ； in, 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 after using the same temperature parameters for students and teachers, Weight coefficient, is the square term of temperature; 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 tasks offloaded from the terminal.

6. The UAV agent-edge-cloud layered decision-making method according to claim 1 is characterized in that: In step S4, the step of performing global optimization decision based on the large-scale pre-trained model specifically includes: S401, construct input data format: structure the complex mission data uploaded by the drone and construct the input sequence , expressed as: ； in, represents the input task feature vector, Encode for position; S402, Transformer encoding layer processing: The input sequence is processed in multiple layers through the Transformer encoder. Each layer includes a multi-head self-attention mechanism and a feedforward network. Attention calculation: ； Then multiple heads perform attention calculations in parallel and then perform splicing processing: ； The output vector of each position is then processed by two layers of linear transformation and activation function through the feedforward network: ； S403. Design pre-training tasks: Pre-train the model based on 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: ； in, , , They represent path optimization loss, image recognition accuracy loss, and delay prediction error loss respectively; S404, decision output in the reasoning phase: In the reasoning phase, a pre-trained Transformer model is used to make global optimization decisions on the computationally intensive or data-intensive tasks forwarded by the edge computing nodes. The output includes but is not limited to path prediction results, target recognition categories, or control strategies.

7. The UAV agent-edge-cloud layered decision-making method according to claim 1 is characterized in that: In step S5, the method for optimizing overall resource allocation is to monitor network bandwidth, latency, and the load status of each computing node in real time based on a dynamic load sensing mechanism, and then dynamically adjust the computing resource allocation ratio between the end, edge, and cloud using a multi-objective optimization algorithm. Specifically, the following steps are included: S501. Construct state space: Construct system state vector , real-time monitoring of network bandwidth , communication delay , and the resource utilization of the three types of computing nodes, namely, end, edge, and cloud, are expressed as: ； in, , , Represents the CPU / GPU resource utilization on the device, edge, and cloud sides respectively; S502. Define action space: define resource scheduling action space , which represents the computing resource allocation ratio of the three types of nodes: end, edge, and cloud, so that it satisfies: ； in, , , The proportion of computing resources allocated to terminals, edges, and clouds respectively; S503. Design reward function: decision delay , energy consumption and task processing accuracy To optimize the goal, construct a multi-objective optimization reward function : ； in, is the weighting coefficient; S504, reinforcement learning decision-making mechanism: Adopting a resource scheduling strategy based on reinforcement learning to optimize the reward function: ； in, is the discount factor, is the policy function; S505, perform resource scheduling: allocate resources according to the output of the policy function , control the distribution and scheduling of computing tasks among the end, edge, and cloud layers, and dynamically optimize the allocation of computing resources.

8. A UAV agent-end-edge-cloud layered decision system, used to implement the UAV agent-end-edge-cloud layered decision method according to any one of claims 1 to 7, characterized in that: include: UAV terminal module: It has a built-in lightweight deep learning model, a model compression module, and a dynamic task offloading module based on the self-attention mechanism. It is used to process urgent or simple tasks in real time and adaptively offload complex tasks based on task characteristics. Edge computing node module: used to deploy a lightweight local decision model based on knowledge distillation, pre-process and make local decisions on terminal offload tasks, and determine whether the tasks should be forwarded to the cloud; Cloud data center module: Equipped with a large-scale pre-trained model based on the Transformer structure, it is used to perform global optimization decisions for compute-intensive or data-intensive tasks forwarded by 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 optimized allocation of computing resources between drone terminals, edge computing nodes and cloud data centers based on multi-objective optimization algorithms.

Images