Methods, devices, electronic equipment and storage media for unmanned aerial vehicle (UAV) mission offloading in low-Earth orbit satellite networks

Abstract

This invention provides a method, apparatus, electronic device, and storage medium for unmanned aerial vehicle (UAV) mission offloading in a low-Earth orbit (LEO) satellite network, relating to the field of communication technology. Based on the LEO satellite network, a joint optimization model is constructed with the objective of maximizing long-term average mission offloading. Virtual computing energy consumption queues and virtual resource wholesale cost queues are constructed to address long-term average computing energy consumption constraints and long-term average cost constraints, respectively. Based on the Lyapunov optimization framework, the joint optimization model is transformed into a single-slot static optimization problem. Within each time slot, based on the current UAV mission arrival status and virtual queue backlog status, the single-slot static optimization problem is solved to obtain the service deployment decision, computing resource allocation decision, and mission offloading decision to the ground network for that time slot. This effectively ensures that the satellite network maintains a high mission offloading rate during long-term continuous operation.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and more specifically, to a method, apparatus, electronic device, and storage medium for unmanned aerial vehicle (UAV) mission offloading in a low-Earth orbit satellite network. Background Technology

[0002] In low-Earth orbit (LEO) satellite-UAV collaborative networks, traditional online task offloading methods typically treat each time slot as an independent time unit. Within each time slot, the task offloading location and resource allocation scheme are determined solely based on the current UAV task arrival volume, the channel status between the satellite and the UAV, and the currently available computing resources of the satellite node. This type of method fails to constrain the computing energy consumption of satellite nodes and the cost of wholesale computing resources from the satellite network to the terrestrial network as long-term indicators that need to be uniformly managed across multiple consecutive time slots. Since LEO satellites rely on limited solar power, if the computing energy consumption of a single satellite remains high for several consecutive time slots, it will lead to a rapid depletion of remaining energy, making it impossible to maintain basic computing capabilities in subsequent time slots. Similarly, the computing power services provided by the terrestrial network are usually billed on a daily / weekly basis or managed by quotas. While wholesale resources in a single time slot can increase the current task processing volume, it can easily cause the average cost over the entire period to exceed the budget limit, triggering service frequency throttling or tariff penalty mechanisms.

[0003] Therefore, traditional methods lack the means to model and control the two key operational constraints of "long-term average computing energy consumption" and "long-term average wholesale cost", making it difficult to guarantee the stable service capability and sustainable operation of satellite networks over long time scales. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method, apparatus, electronic device and storage medium for unmanned aerial vehicle (UAV) mission offloading in a low-Earth orbit satellite network to at least partially improve the above-mentioned problems.

[0005] To achieve the above objectives, the technical solutions adopted in the embodiments of the present invention are as follows: In a first aspect, embodiments of the present invention provide a method for unmanned aerial vehicle (UAV) mission offloading in a low-Earth orbit (LEO) satellite network. The LEO satellite network includes multiple LEO satellites, multiple UAVs, and a ground network. Each LEO satellite is interconnected via an inter-satellite link and is communicatively connected to both the UAVs and the ground network. The method includes: Based on the aforementioned low-Earth orbit satellite network, a joint optimization model is constructed with the objective of maximizing the long-term average task offload. The joint optimization model includes the long-term average computing energy consumption constraint of the low-Earth orbit satellites and the long-term average cost constraint of the low-Earth orbit satellite network wholesaling computing resources to the ground network. A virtual computing energy consumption queue and a virtual resource wholesale cost queue are constructed for the long-term average computing energy consumption constraint and the long-term average cost constraint, respectively. Based on the Lyapunov optimization framework, the joint optimization model is transformed into a single-slot static optimization problem; Within each time slot, based on the arrival status of UAV missions and the backlog status of the virtual queue in the current time slot, the single-time slot static optimization problem is solved to obtain the service deployment decision, computing resource allocation decision, and mission offloading decision to the ground network for that time slot.

[0006] Optionally, constructing a virtual computing energy consumption queue and a virtual resource wholesale cost queue for the long-term average computing energy consumption constraint and the long-term average cost constraint respectively includes: The virtual computational energy consumption queue for each low-Earth orbit satellite is updated according to the following formula:

[0007] in, Indicates low-orbit satellites In the time slot The backlog length of the virtual computing energy consumption queue. For low-orbit satellites The long-term average energy consumption threshold, For low-orbit satellites For drones Type Computing resources allocated to the task The energy consumption coefficient corresponding to the task performed by the satellite unit. This is the energy consumption factor amplification factor. The length of each time slot, For low-Earth orbit satellites, For time slots A collection of drones covered by an inner low-Earth orbit satellite network; Update the virtual resource wholesale cost queue for the entire low-Earth orbit satellite network according to the following formula:

[0008] in, For time slots The backlog length of the virtual wholesale cost queue. The long-term average cost threshold. Unmanned aerial vehicle (UAV) terminals for offloading from low-Earth orbit (LEO) satellite networks to terrestrial networks The workload, For drone terminals In the time slot Chinese type The total arrival requirements of the task. Calculate the unit resource price for the terrestrial network. This is the wholesale resource cost ratio coefficient.

[0009] Optionally, the transformation of the joint optimization model into a single-slot static optimization problem based on the Lyapunov optimization framework includes: The Lyapunov function is defined as half the sum of the squares of all the virtual computing energy consumption queues and the virtual resource wholesale cost queues; The single-slot Lyapunov offset function is defined as the expected increment of the Lyapunov function of the current slot relative to the previous slot; The Lyapunov offset plus penalty function is constructed as the single-slot Lyapunov drift minus the product of the Lyapunov weight coefficient and the expected total task unloading amount in the current slot; By minimizing the upper bound of the Lyapunov offset plus penalty function, the joint optimization model is transformed into a single-slot static optimization problem that depends only on the current slot state and the virtual queue backlog in each slot.

[0010] Optionally, within each time slot, based on the arrival status of UAV missions and the backlog status of the virtual queue in the current time slot, the single-time-slot static optimization problem is solved to obtain the service deployment decision, computing resource allocation decision, and mission offloading decision to the ground network for that time slot, including: The arrival status of drone missions and the backlog status of virtual queues in the current time slot are input into a pre-trained deep neural network to obtain relaxed service deployment decisions. The relaxed service deployment decision is extended to a set of candidate service deployment decisions that satisfy the satellite storage capacity constraint; For each of the candidate service deployment decisions, after fixing the service deployment decision, solve the subproblems involving only continuous variables in the single time slot static optimization problem to obtain the target reward value and the corresponding task unloading decision and computing resource allocation decision; The candidate service deployment decision with the largest target reward value, along with its corresponding task unloading decision and computing resource allocation decision, is selected as the final decision for the current time slot.

[0011] Optionally, extending the relaxed service deployment decision to a set of candidate service deployment decisions that satisfy satellite storage capacity constraints includes: The relaxed service deployment decision is divided into multiple segments according to a preset method to obtain multiple candidate action sub-vectors; wherein, the length of each candidate action sub-vector is the number of service types, and each candidate action sub-vector corresponds to the relaxation value of all service types on a low-Earth orbit satellite; For each candidate action subvector, an initial feasible service deployment decision is generated using a greedy strategy. An enhancement approach based on K-nearest neighbors is adopted to generate multiple new feasible service deployment decisions based on the initial feasible service deployment decision; All generated feasible service deployment decisions are used as candidate service deployment decisions.

[0012] Optionally, the K-nearest neighbor-based enhancement method generates multiple new feasible service deployment decisions based on the initial feasible service deployment decision, including: Based on the initial feasible service deployment decision, initialize the set of deployed services, the set of undeployed services, and the remaining capacity; Construct a first sorting queue and a second sorting queue; wherein, the first sorting queue is formed by calculating the absolute value of the relaxation value difference between each deployed service and each undeployed service, and sorting them in ascending order of the absolute value, to represent the nearest neighbor relationship between deployed services and undeployed services; the second sorting queue is formed by calculating the absolute value of the relaxation value difference between each undeployed service and each deployed service, and sorting them in ascending order of the absolute value, to represent the nearest neighbor relationship between undeployed services and deployed services. Set a preset number of neighbors K, select the first K deployed services from the first sorting queue, and select the first K undeployed services from the second sorting queue; For each selected deployed service and each undeployed service, exchange their deployment states and generate a new service deployment decision; Each time a new service deployment decision is generated, it is checked whether it meets the satellite storage capacity constraint. If it does, it is considered a new feasible service deployment decision.

[0013] Optionally, the method further includes: Store the current state obtained in each time slot and the selected optimal service deployment decision as a sample; Every preset period, a batch of samples is randomly selected, and the parameters of the deep neural network are updated using the Adam algorithm.

[0014] Secondly, embodiments of the present invention provide a drone mission offloading device for a low-Earth orbit (LEO) satellite network. The LEO satellite network includes multiple LEO satellites, multiple drones, and a ground network. Each LEO satellite is interconnected via an inter-satellite link and is communicatively connected to both the drones and the ground network. The device includes: The model building unit is used to construct a joint optimization model based on the low-Earth orbit satellite network with the objective of maximizing the long-term average task offload. The joint optimization model includes the long-term average computing energy consumption constraint of the low-Earth orbit satellite and the long-term average cost constraint of the low-Earth orbit satellite network to wholesale computing resources to the ground network. The queue construction unit is used to construct a virtual computing energy consumption queue and a virtual resource wholesale cost queue for the long-term average computing energy consumption constraint and the long-term average cost constraint, respectively. The virtual computing energy consumption queue represents the cumulative deviation of the actual computing energy consumption of each of the low-Earth orbit satellites from its long-term average energy consumption constraint in each time slot, and the virtual resource wholesale cost queue represents the cumulative deviation of the actual resource wholesale cost of the low-Earth orbit satellite network from its long-term average cost constraint in each time slot. The model transformation unit is used to transform the joint optimization model into a single-slot static optimization problem based on the Lyapunov optimization framework. The transformation is achieved by minimizing the single-slot drift plus penalty function, which includes the task unloading benefit term of the current slot, the penalty term for virtual queue backlog, and Lyapunov weight coefficients used to adjust the trade-off between task processing performance and the degree of long-term constraint satisfaction. The problem-solving unit is used to solve the single-time-slot static optimization problem in each time slot based on the arrival status of UAV tasks and the backlog status of the virtual queue in the current time slot, and obtain the service deployment decision, computing resource allocation decision and task offloading decision to the ground network for that time slot.

[0015] Thirdly, embodiments of the present invention provide an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the method described in any of the above-mentioned embodiments.

[0016] Fourthly, embodiments of the present invention provide a storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in any of the preceding claims.

[0017] This invention provides a method, apparatus, electronic device, and storage medium for unmanned aerial vehicle (UAV) mission offloading in a low-Earth orbit (LEO) satellite network. By constructing a virtual computing energy consumption queue and a virtual resource wholesale cost queue, and based on the Lyapunov optimization framework, the long-term average constraints are explicitly transformed into local optimization objectives for each time slot. This achieves online solvable collaborative decision-making for a single time slot without relying on prior information about future missions, effectively ensuring that the satellite network maintains a high mission offloading rate while meeting the constraints of long-term average energy consumption and wholesale cost during long-term continuous operation.

[0018] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A schematic structural block diagram of an electronic device provided in an embodiment of the present invention; Figure 2 A schematic structural block diagram of a low-Earth orbit satellite network provided in an embodiment of the present invention; Figure 3 A flowchart illustrating a method for unmanned aerial vehicle (UAV) mission offloading in a low-Earth orbit satellite network, provided as an embodiment of the present invention; Figure 4 A flowchart illustrating step S230 provided in an embodiment of the present invention; Figure 5 A flowchart illustrating step S240 provided in an embodiment of the present invention; Figure 6 A flowchart of a Ly-DRTOSD provided as an embodiment of the present invention; Figure 7 This is a schematic structural block diagram of an unmanned aerial vehicle (UAV) mission unloading device for a low-Earth orbit satellite network, provided as an embodiment of the present invention.

[0021] Icons: 100-Electronic device; 101-Memory; 102-Communication interface; 103-Processor; 104-Communication bus; 300-Unmanned aerial vehicle mission offloading device for low-orbit satellite network; 310-Model building unit; 320-Queue building unit; 330-Model conversion unit; 340-Problem solving unit. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0023] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0024] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0025] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0026] To implement the process steps and functions of this invention, please refer to [link / reference]. Figure 1 , Figure 1 This is a schematic structural block diagram of an electronic device 100 provided in an embodiment of the present invention. The electronic device 100 includes a memory 101 and a processor 103, which are electrically connected directly or indirectly to each other to realize data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses 104 or signal lines. The memory 101 can be used to store software programs and modules, and the processor 103 executes the software programs and modules stored in the memory 101, thereby performing various functional applications and data processing.

[0027] Electronic device 100 can be, but is not limited to, a personal computer (PC), a server, a distributed computer, etc. It is understood that electronic device 100 is not limited to a physical server, but can also be a virtual machine on a physical server, a virtual machine built on a cloud platform, or any other computer that can provide the same functionality as the server or virtual machine. The operating system of electronic device 100 can be, but is not limited to, Windows, Linux, etc.

[0028] The memory 101 may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0029] The communication connection between the electronic device 100 and external devices is achieved through at least one communication interface 102 (which can be wired or wireless).

[0030] Processor 103 may be an integrated circuit chip with signal processing capabilities. In implementation, the steps of this embodiment can be completed by integrated logic circuits in the hardware of processor 103 or by instructions in software form. Processor 103 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0031] Understandable. Figure 1 The structure shown is for illustrative purposes only; the electronic device 100 may also include components that are more advanced than those shown. Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown. Figure 1 The components shown can be implemented using hardware, software, or a combination thereof.

[0032] The following is an exemplary description of the unmanned aerial vehicle (UAV) mission offloading method for low-Earth orbit satellite networks provided in an embodiment of the present invention. See also... Figure 2 The low-Earth orbit (LEO) satellite network consists of multiple LEO satellites, multiple drones, and a ground network. The LEO satellites are interconnected via inter-satellite links and communicate with the drones and the ground network respectively.

[0033] See Figure 3 The subject executing this method can be one of the above. Figure 1 The electronic device 100 shown, the method includes as follows Figure 3 The following steps are described: S210: Based on the low-Earth orbit satellite network, construct a joint optimization model with the objective of maximizing the long-term average mission unloading.

[0034] The joint optimization model includes the long-term average computing energy consumption constraint of low-Earth orbit satellites and the long-term average cost constraint of low-Earth orbit satellite networks wholesaling computing resources to ground networks.

[0035] S220: Construct a virtual computing energy consumption queue and a virtual resource wholesale cost queue for the long-term average computing energy consumption constraint and the long-term average cost constraint, respectively.

[0036] Among them, the virtual computing energy consumption queue represents the cumulative deviation of the actual computing energy consumption of each low-Earth orbit satellite in each time slot from its long-term average energy consumption constraint, while the virtual resource wholesale cost queue represents the cumulative deviation of the actual resource wholesale cost of the low-Earth orbit satellite network in each time slot from its long-term average cost constraint.

[0037] S230: Based on the Lyapunov optimization framework, the joint optimization model is transformed into a single-slot static optimization problem.

[0038] The transformation is achieved by minimizing a single-slot drift plus penalty function, which includes the task unloading benefit term for the current slot, the penalty term for virtual queue backlog, and Lyapunov weight coefficients used to adjust the trade-off between task processing performance and the degree of long-term constraint satisfaction.

[0039] S240: Within each time slot, based on the arrival status of UAV missions and the backlog status of the virtual queue in the current time slot, solve the single-time slot static optimization problem to obtain the service deployment decision, computing resource allocation decision, and mission offloading decision to the ground network for that time slot.

[0040] This method constructs a joint optimization model incorporating long-term average computing energy consumption constraints and long-term average resource wholesale cost constraints. It dynamically quantifies these two long-term constraints using virtual computing energy consumption queues and virtual resource wholesale cost queues. Then, based on the Lyapunov optimization framework, the cross-time-slot coupled long-term objective is transformed into a static optimization problem solvable within each time slot solely by the current task arrival status and queue backlog status. Ultimately, without predicting future tasks, it achieves integrated online collaborative decision-making for service deployment, resource allocation, and task offloading within a three-tiered architecture of satellite, UAV, and terrestrial network. This overcomes the problems of satellite endurance degradation and uncontrolled terrestrial costs caused by neglecting long-term operational constraints in traditional methods. While ensuring long-term stable system operation, it continuously improves task offloading efficiency and overall network service capabilities.

[0041] The low-Earth orbit (LEO) satellite network consists of multiple LEO satellites, multiple UAVs, and a ground network. The LEO satellites are interconnected through inter-satellite links to form an on-orbit collaborative computing plane. Each UAV accesses the LEO satellite covering its current location through a wireless link (e.g., Ka-band). Each LEO satellite communicates with the ground network through a satellite-to-ground link (e.g., laser or microwave link), thus forming a three-level collaborative architecture of "UAV-LEO satellite-ground network".

[0042] Based on the aforementioned network architecture, the system optimization objective is set to maximize the long-term average task offloading over a given time span. This joint optimization model not only considers the task offloading decision itself but also incorporates two types of long-term operational constraints: first, the long-term average computing energy consumption of each low-Earth orbit satellite must not exceed its preset energy consumption threshold to ensure its continuous on-orbit power supply capability; second, the long-term average cost generated by the entire low-Earth orbit satellite network in wholesaling computing resources to the ground network must not exceed its budget limit to maintain the economic sustainability of satellite-ground collaborative services.

[0043] In one alternative implementation, for ease of description, the entire time span is divided into equal-length time slots, with the length of each time slot denoted as . Within each time slot, the satellite network traverses multiple geographical regions, where unmanned aerial vehicle (UAV) terminals are distributed to perform tasks such as inspection and monitoring, environmental perception, disaster identification, target tracking, and emergency communication. During flight, UAVs continuously generate task requests characterized by real-time processing and computational intensity. The satellite network completes as many of these tasks as possible through local processing or inter-satellite collaborative processing, thus constituting an online collaborative offloading scenario oriented towards task flow.

[0044] Assuming in time slot In China, there are a total of The drone was connected to the satellite network and requested... Different types of task processing services. For symbolic representation, time slot sets, low-Earth orbit satellite sets, and time slots are respectively... The set of drones and the set of service types that have access to the satellite network are denoted as , , and .

[0045] Due to the limited storage and computing resources of satellite nodes, only a limited number of service types can be deployed within the same time slot of any given satellite. This means that some services are active and can be directly executed by the onboard processor, while others are not loaded and cannot provide processing capabilities directly within the current time slot. For a specific type of UAV mission, the target satellite can only receive and process the mission if the corresponding service instance has been pre-deployed; otherwise, the mission cannot be directly executed by the current satellite but must be forwarded via inter-satellite links to other satellite nodes that have deployed the corresponding service for collaborative processing.

[0046] Therefore, in the network scenario considered in this embodiment of the invention, task execution capability depends not only on the connection between the UAV and the access satellite, but also on the service deployment status of each satellite node, available computing resources, and the collaborative forwarding capability of the inter-satellite links. Based on this, this embodiment of the invention aims to maximize the long-term average task offload of the system, comprehensively considering satellite service deployment, task offload decisions, and computing resource allocation, and further taking into account the long-term computing endurance of satellite nodes and the resource overhead introduced by the collaborative forwarding of tasks via inter-satellite links, thereby achieving efficient support for dynamic UAV task flows.

[0047] Low-Earth orbit satellites communicate via laser links. Without loss of generality, the communication latency between satellites is negligible due to the high data transmission rate of laser communication. Unmanned aerial vehicle (UAV) terminals establish connections with satellites via the Ka-band spectrum, and the satellites transmit data to ground network infrastructure via laser links. Indicates the drone terminal in the time slot The available bandwidth, considering large-scale attenuation and shadow-Rice fading, will be used to power the drone terminal. With satellite The channel gain between them is denoted as The upload data rate of the drone terminal can then be expressed as:

[0048] in, Indicates drone terminal Data upload power, This indicates environmental noise.

[0049] In the application scenario described in this embodiment of the invention, the drone terminal continuously generates task flows, and tasks can be processed in parallel within a time slot. The time slot... Internal source: drone terminal Type Task arrival rate is expressed as It indicates the type of task completed per unit of time. The computational load required for the task remains constant throughout the time slot. Based on the above definition, the UAV terminal can be further obtained. In the time slot Chinese type The total arrival requirement of the task is denoted as .

[0050] Assume that the satellite network and edge network process the tasks in real time and return the results. In the above-mentioned task collaborative offloading scenario, the satellite collaboratively allocates computing resources based on the task arrival rate of each UAV terminal, and offloads some tasks to the ground network, using a parallel task processing mechanism to achieve real-time offloading of the UAV task flow.

[0051] Considering the limited local computing resources of UAV terminals, this invention leverages the task processing capabilities of low-Earth orbit (LEO) satellite networks. By combining inter-satellite and satellite-to-ground collaboration mechanisms, it enables the satellite network and terrestrial network to execute as many tasks generated by the UAV terminals as possible. Based on the aforementioned network scenario, for any type of task, LEO satellites must pre-deploy service instances corresponding to that task type before executing it.

[0052] Let the binary variable represent satellite In the time slot Internal deployment type The service, when the variable takes a value of 1, indicates satellite In the time slot The corresponding service is deployed internally; when the variable takes a value of 0, it indicates that the corresponding service is not deployed. Due to the limited storage capacity of satellite nodes, only a limited number of service types can be deployed simultaneously in the same time slot. Therefore, the following constraint inequality must be satisfied:

[0053] in, Indicates running service Required server resources Indicates satellite Total capacity.

[0054] Furthermore, using Indicates satellite In the time slot The interior contains the drone terminal. Allocated computing resources. Since the total allocable computing resources to a satellite server cannot exceed its maximum computing capacity, Indicates low-orbit satellites Given the maximum computing resources, the following constraint inequality holds:

[0055] Low-Earth orbit (LEO) satellites perform computational tasks by consuming their own energy, which impacts the continuous operational capability of the satellite nodes. Therefore, the average computational energy consumption of the satellite during long-term operation should be controlled within a given range. Simultaneously, considering that LEO satellites are also subject to instantaneous maximum computational energy consumption limitations within a single time slot, the long-term and instantaneous constraints on LEO satellite computational energy consumption can be further derived, as follows:

[0056]

[0057] The first constraint represents the maximum instantaneous energy consumption limit within a single time slot, while the second constraint represents the long-term average computational energy consumption limit. Indicates satellite The energy consumption coefficient corresponding to the task being performed by the executing unit. This represents the energy consumption factor amplification factor, used to control the components of the objective function and constraints to be in the same dimension; This represents the instantaneous maximum computational energy consumption of a single time slot. This represents the long-term average energy consumption threshold.

[0058] In addition, satellite networks can offload some tasks to ground networks, allowing for the offloading of more tasks through a satellite-ground collaborative approach. This indicates the drone terminal that offloads data from the satellite network to the terrestrial network. The workload. Meanwhile, the satellite network's request for computing resources from the terrestrial network incurs additional wholesale resource costs, especially when the satellite and terrestrial networks are not operated by the same carrier; this cost is significant. Therefore, time slots... The total wholesale resource cost can be expressed as:

[0059] in, This represents the unit computing resource price of the terrestrial network, and... similar, This represents the wholesale resource cost ratio coefficient.

[0060] Therefore, drone terminals In the time slot Internal correspondence to type The total amount of task unloading can be expressed as:

[0061] In this embodiment of the invention, satellite networks and terrestrial networks collaborate to provide task offloading services for UAV terminals, with the optimization objective being to maximize the system's long-term average task offloading volume. However, this optimization objective may cause low-Earth orbit satellites to consume more computing energy to perform tasks, or prompt the satellite network to request more computing resources from the terrestrial network, resulting in higher wholesale resource costs. This will pose challenges to maintaining the satellite's computing endurance and controlling the wholesale resource costs of the terrestrial network.

[0062] Therefore, the joint optimization problem of the embodiments of the present invention can be formalized into the following optimization problem:

[0063] Among them, constraint C5 indicates that the long-term average wholesale resource cost generated by the satellite network cannot exceed a given threshold. Constraint C6 states that the total task offloading amount cannot exceed the task demand in each time slot; constraint C7 states the range of values ​​for computing resource allocation. Furthermore, constraint C7 also restricts the following relationship: if a satellite wants to allocate computing resources for a certain type of task, it must first deploy a service instance corresponding to that type of task.

[0064] Furthermore, due to the aforementioned optimization problem The decision variables in the problem are coupled with each other across different time slots. At the same time, there are also significant coupling relationships between task unloading decisions, computing resource allocation decisions, and service deployment decisions. Therefore, this problem is difficult to solve.

[0065] To reduce the complexity of solving the original problem, this embodiment of the invention first uses the Lyapunov optimization framework to decouple the original long-term optimization problem by time slot, and then transforms it into a single-time slot optimization problem. On this basis, a two-stage optimization method based on deep reinforcement learning is then used to solve the decoupled single-time slot problem.

[0066] After the joint optimization model is constructed, step S220 is executed. For each low-Earth orbit (LEO) satellite, a separate virtual computational energy consumption queue is constructed based on its long-term average computational energy consumption constraint. This queue represents the cumulative deviation between the actual computational energy consumption of the satellite and its long-term average energy consumption threshold across historical time slots. For the entire LEO satellite network, a global virtual resource wholesale cost queue is constructed based on its long-term average wholesale cost constraint. This queue represents the cumulative deviation between the network's actual wholesale cost and its long-term average cost threshold across historical time slots. Both virtual queues are dynamically updated over time, and their current backlog length directly reflects the risk level of the corresponding long-term constraint being violated.

[0067] In one alternative implementation, since service deployment decisions, task offloading decisions, and computing resource allocation decisions all change dynamically over time slots, the computing energy consumption and wholesale cost of ground resources for the satellite network may be higher or lower than their respective long-term average thresholds in different time slots. Therefore, step S220 may include: The virtual computational energy consumption queue for each low-Earth orbit satellite is updated according to the following formula:

[0068] in, Indicates low-orbit satellites In the time slot The backlog length of the virtual computing energy consumption queue. For low-orbit satellites The long-term average energy consumption threshold, For low-orbit satellites For drones Type Computing resources allocated to the task The energy consumption coefficient corresponding to the task performed by the satellite unit. This is the energy consumption factor amplification factor. The length of each time slot, For low-Earth orbit satellites, For time slots A collection of drones covered by an inner low-Earth orbit satellite network.

[0069] Update the virtual resource wholesale cost queue for the entire low-Earth orbit satellite network according to the following formula:

[0070] in, For time slots The backlog length of the virtual wholesale cost queue. The long-term average cost threshold. Unmanned aerial vehicle (UAV) terminals for offloading from low-Earth orbit (LEO) satellite networks to terrestrial networks The workload, For drone terminals In the time slot Chinese type The total arrival requirements of the task. Calculate the unit resource price for the terrestrial network. This is the wholesale resource cost ratio coefficient.

[0071] Next, a single-slot static optimization problem is transformed. Using Lyapunov optimization theory, the original cross-slot coupled long-term optimization problem is transformed into a series of independent single-slot optimization problems. This transformation is achieved by constructing and minimizing a single-slot drift plus penalty function, which consists of three parts: (1) a task offloading benefit term available in the current slot, reflecting short-term performance; (2) a penalty term positively correlated with the current virtual queue backlog length, reflecting suppression of long-term constraint deviation; and (3) an adjustable Lyapunov weight coefficient, used for flexible trade-offs between task processing performance and the degree of long-term constraint satisfaction. Optionally, see [link to relevant documentation]. Figure 4 Step S230 may include: S231: The Lyapunov function is defined as half the sum of the squares of all virtual computing energy consumption queues and virtual resource wholesale cost queues.

[0072] S232: Define the single-slot Lyapunov offset function as the expected increment of the Lyapunov function of the current slot relative to the previous slot.

[0073] S233: Construct the Lyapunov offset plus penalty function as the product of the single-slot Lyapunov drift minus the Lyapunov weight coefficient and the expected total task unloading amount in the current slot.

[0074] S234: By minimizing the upper bound of the Lyapunov offset plus penalty function, the joint optimization model is transformed into a single-slot static optimization problem that depends only on the current slot state and the virtual queue backlog in each slot.

[0075] Will and Initialize to 0, controlled by The stability of the constraint can guarantee the conditions C4 and C5.

[0076] Next, we introduce the Lyapunov function as follows:

[0077] Furthermore, the Lyapunov offset function can be defined as:

[0078] To ensure the virtual queue To achieve stability while maximizing the long-term average task offload, this embodiment of the invention employs an offset penalty function minimization method, constructing the Lyapunov offset penalty function as follows:

[0079] in, This parameter represents the trade-off between task offloading and queue backlog. By introducing this parameter, the original long-term optimization objective can be transformed into a parameter for each time slot. Internal relative to the corresponding upper bound Minimize.

[0080] According to Lyapunov's upper bound theorem on drift 1: for each time slot All of them have the following Upper bound:

[0081] in, It is a constant, and .

[0082] According to this theorem, each time slot can be obtained. of The upper limit, the original problem is transformed as follows:

[0083] Through the above steps, the original long-running optimization problem is solved. It is transformed into a single-slot static optimization problem .

[0084] Due to the problems after the transformation This problem still falls under the category of mixed-integer nonlinear programming and retains high solution complexity. Therefore, this invention further designs a two-stage online optimization method based on deep reinforcement learning to improve the solution efficiency of the single-slot problem. See also... Figure 5 Step S240 may include: S241: Input the arrival status of the drone mission and the virtual queue backlog status of the current time slot into a pre-trained deep neural network to obtain a relaxed service deployment decision.

[0085] S242: Extend the relaxed service deployment decision to a set of candidate service deployment decisions that meet the satellite storage capacity constraint.

[0086] S243: For each candidate service deployment decision, after fixing the service deployment decision, solve the subproblems involving only continuous variables in the single-slot static optimization problem to obtain the target reward value and the corresponding task unloading decision and computing resource allocation decision.

[0087] S244: Select the candidate service deployment decision with the largest target reward value and its corresponding task unloading decision and computing resource allocation decision as the final decision for the current time slot.

[0088] The single-slot optimization problem after decoupling Analysis shows that it occurs in time slots Decision variables within It mainly includes three categories, namely binary service deployment decision variables. Continuous task unloading decision variables and continuous resource allocation decision variables Since the complexity of solving this problem mainly stems from the discrete service deployment decision variables and their coupling relationship with task unloading and resource allocation variables, this embodiment of the invention employs a two-stage approach to solve the single-slot problem. Solve the problem.

[0089] When the deployment variables for binary services are fixed, the problem arises. This can be transformed into a linear programming problem involving only continuous variables, thus enabling a solution in polynomial time using methods such as the interior-point method. Therefore, in time slots... Inside, can be used The network state is constructed by representing the current task arrival volume and virtual queue backlog status of the drone terminal; for specific service deployment variables... ,use This indicates that by optimizing variables and The optimal value was obtained. At this point, the problem... Solving this problem is equivalent to finding the optimal service deployment. ,Right now:

[0090] question The goal is to obtain a strategy After sufficient training, it can respond to input Rapidly generate optimal service deployment decisions, i.e. .

[0091] In the time slot At the beginning, As the current state of the network, input to the parameter... A fully connected neural network (DNN). The DNN will adjust its behavior based on the current state and strategy. Generate relaxed service deployment decisions .

[0092] Next, execute S242, according to Generate a set of service deployment decisions The service deployment decision can be generated using the Order-preserving quantization (OPQ) method. However, for problems with constraints on actions, the OPQ method cannot guarantee the feasibility of each generated action, which can adversely affect the convergence of the training process. To overcome this problem, this embodiment of the invention proposes an improved K-nearest neighbor enhancement method based on the OPQ method. Through a two-stage process of greedy initialization and K-nearest neighbor enhancement, the relaxed actions are expanded into a set of feasible service deployment decisions that satisfy satellite capacity constraints and have better discreteness, thereby improving the convergence of training and the feasibility of actions. In an optional implementation, step S242 may include: S2421: Divide the relaxed service deployment decision into multiple segments according to a preset method to obtain multiple candidate action sub-vectors.

[0093] The length of each candidate action subvector is the number of service types, and each candidate action subvector corresponds to the relaxation value of all service types on a low-Earth orbit satellite.

[0094] S2422: For each candidate action subvector, an initial feasible service deployment decision is generated using a greedy strategy.

[0095] S2423: Employs a K-nearest neighbor-based enhancement approach to generate multiple new feasible service deployment decisions based on the initial feasible service deployment decision.

[0096] S2423 may include: S24231: Based on the initial feasible service deployment decision, initialize the set of deployed services, the set of undeployed services, and the remaining capacity.

[0097] S24232: Construct the first sorting queue and the second sorting queue.

[0098] The first sorting queue is used to calculate the absolute value of the relaxation value difference between each deployed service and each undeployed service, and sort them in ascending order of the absolute value, which is used to characterize the nearest neighbor relationship between deployed and undeployed services; the second sorting queue is used to calculate the absolute value of the relaxation value difference between each undeployed service and each deployed service, and sort them in ascending order of the absolute value, which is used to characterize the nearest neighbor relationship between undeployed and deployed services.

[0099] S24233: Set a preset number of neighbors K, select the first K deployed services from the first sorting queue, and select the first K undeployed services from the second sorting queue.

[0100] S24234: For each selected deployed service and each undeployed service, exchange their deployment states and generate a new service deployment decision.

[0101] S24235: For each new service deployment decision generated, check whether it meets the satellite storage capacity constraint. If it does, it is considered a new feasible service deployment decision.

[0102] S2424: Treat all generated feasible service deployment decisions as candidate service deployment decisions.

[0103] See Figure 6 The present invention embodiment can refer to this method as a two-stage online service deployment based on Lyapunov optimization and deep reinforcement learning (Lyapunov-based Deep Reinforcement Learning Two-stage Online Service Deployment, Ly-DRTOSD), which will be described below.

[0104] In this context, the state space of Ly-DRTOSD represents the input of the DNN, consisting of the UAV mission arrival state and the virtual queue backlog state in the current time slot. Therefore, the time slot... The state can be represented as:

[0105] Ly-DRTOSD's action space represents service deployment decisions, and can be used... In Ly-DRTOSD, the DNN first generates a lenient service deployment decision based on the current time slot's policy and state. For training to proceed, this relaxed service deployment decision needs to be extended to a set of binary service deployment decisions.

[0106] According to convex optimization theory, Lemma 1 states that any optimization problem... ,in, It is a convex function. It is a convex set. This is the corresponding optimal solution for any convex set. ,if , ,and If it is the optimal solution, then we have .

[0107] It can be seen that deploying as many services as possible within the constraints will not lead to a performance degradation.

[0108] Theorem 2: ,in Theorem 2 shows that as many services as possible should be deployed within the constraints, and therefore a greedy strategy can be used to achieve this. Generate the first service deployment decision, and based on this, generate a set of feasible service deployment decisions.

[0109] First, execute S2421 to divide the relaxed service deployment vector output by the neural network in the current time slot into a preset format. Segments, each segment's length equals the number of service types. Each candidate action subvector is treated as a separate object to be enhanced.

[0110] The upcoming relaxation service deployment vector will be evenly divided according to the number of satellites. For example, if there are eight low-Earth orbit satellites in the current network, the entire list will be divided into eight segments; each segment corresponds to one satellite, and the number of values ​​in each segment equals the total number of service types supported by the system. In this way, each segment represents the degree of deployment preference of the corresponding satellite for various services, forming an independent candidate action sub-vector.

[0111] Subsequently, S2422 and S2423 are executed to perform a two-stage action generation process for each candidate action subvector.

[0112] In the first phase, a greedy strategy is used to generate the first feasible service deployment action. First, the deployment vector corresponding to the current candidate action is initialized, the deployment status of each service type is set to undeployed, and the current relaxed action sub-vector is retained as a set to be processed. Then, provided the set to be processed is not empty, the service type with the largest current value is repeatedly selected as the priority deployment target. If the current satellite's remaining capacity can accommodate the service, the corresponding service deployment status is set to deployed, and the set to be processed and the remaining capacity are updated synchronously. If the current remaining capacity is insufficient to accommodate the service, deployment is not performed, and the service is simply removed from the set to be processed. This process is repeated until the set to be processed is empty, thus obtaining the first service deployment action that satisfies the capacity constraint, and this action is added to the enhancement action set.

[0113] In the second stage, an enhancement method based on K-nearest neighbors is used to further generate other feasible service deployment actions. Based on the feasible deployment actions obtained in the first stage, the set of deployed services, the set of undeployed services, and the remaining capacity are initialized. Then, corresponding sorting queues are constructed, one queue representing the distance relationship between deployed services and undeployed states, and the other queue representing the distance relationship between undeployed services and deployed states. Based on this, the two sorting queues are traversed, and the element values ​​in the queues are swapped in a loop to perturb the current service deployment actions, thereby generating new feasible service deployment actions. Each new action that satisfies the constraints is added to the enhanced action set. After repeating the above two-stage processing on each candidate action subvector, the enhanced action set is obtained and output as the candidate action set for the current time slot service deployment decision.

[0114] Finally, the initial deployment plans generated by each satellite, along with all feasible new plans obtained through the exchange operation, are aggregated to form a complete set of candidate service deployment decisions for the current time slot. Each plan in this set is a physically executable deployment option that meets the hard constraints of hardware resources and is significantly different from one another, and can be directly used for the next step of task offloading and resource allocation evaluation.

[0115] After generating the candidate service deployment decisions, the evaluation and selection phase begins. This involves quantitatively comparing all feasible options to select the set of decisions with the best overall performance for the current time slot. This process consists of two closely linked operations: S243, which performs an independent performance evaluation on each candidate option; and S244, which selects the final execution plan based on the evaluation results.

[0116] In Ly-DRTOSD, the optimal action is achieved by solving... To choose. Defined as a reward function, it can be calculated using the following formula.

[0117]

[0118] in, It is a constant used to penalize deployment actions when constraint C1 is not satisfied. The question structure is as follows:

[0119] This is a linear programming problem involving only continuous variables, which can be solved in polynomial time using tools such as the interior-point method or CPLEX. and .

[0120] After evaluating all candidate solutions one by one, the target reward values ​​of each solution are compared horizontally, and the one with the highest value is selected. The complete decision combination corresponding to this solution, namely its service deployment arrangement, the task offloading path obtained by S243, and the specific computing resource allocation ratio, together constitute the final execution instruction of the current time slot.

[0121] Finally, by solving the problem In To obtain the optimal service deployment decision and the corresponding optimal value The method then stores the result in memory. Therefore, the method may also include: The current state obtained in each time slot and the selected optimal service deployment decision are stored as samples.

[0122] Every preset period, a batch of samples is randomly selected, and the parameters of the deep neural network are updated using the Adam algorithm.

[0123] Every certain time slot Randomly select a batch of samples from memory To train the DNN, where This represents the time slot index set of the selected samples, and therefore the service deployment strategy of the DNN. It will be continuously updated and improved. The Adam algorithm is used to reduce the average cross-entropy loss function, which is defined as:

[0124] in, express Size, This indicates that the DNN is in the time slot. The embedding parameters in the dataset. After training is complete, the DNN parameters are updated in the next time slot. .

[0125] The convergence characteristics of Ly-DRTOSD are analyzed below. First, a static stochastic strategy is introduced. Among them, control decisions (i.e., task unloading, resource allocation, and service deployment decisions) are all in different time slots. Then we have the following lemma 2.

[0126] Lemma 2: If If it is feasible, then for any All of them satisfy the following conditions Strategy :

[0127] in, , yes The optimal objective.

[0128] Based on the lemma, the performance conclusions of the method proposed in the embodiments of the present invention can be further obtained as follows.

[0129] Theorem 3: If the problem If it is feasible, then the following proposition holds: Proposition 1: The average task unloading volume satisfies:

[0130] in, It is a constant representing the solution of a single-slot problem. The distance between the current value and the optimal value.

[0131] Proposition 2: and It is stable.

[0132] Proposition 3: For any value and , Strategy The following inequalities must be satisfied:

[0133] Therefore, the average total queue length satisfies:

[0134] Proof: Because It is feasible, therefore Lemma 2 and Strategy To prove Theorem 1, we have the following inequality:

[0135] Among them, inequalities It can be derived from the formula It is derived that, let Then we have:

[0136] right from arrive Summing both sides, we get:

[0137] because , ,make Proposition 1 is proved.

[0138] Next, skipping proposition 2, let's prove proposition 3 first. Strategy Introducing inequalities Then we have:

[0139] By applying the iterative expectation law, From both sides arrive Summing them up, we get:

[0140] make and combined We can obtain:

[0141] Thus, Proposition 3 is proved. Based on this, we can proceed from... get:

[0142] That is, all energy consumption and wholesale resource cost queues are stable, thus Proposition 2 is proved.

[0143] Theorem 3 shows that the method proposed in this invention has good convergence; at the same time, the objective function of average task unloading, the backlog length of the low-orbit satellite computational energy consumption queue, and the backlog length of the ground wholesale resource cost queue are all related to the Lyapunov weighting coefficients. Related, therefore, the weighting parameters can be adjusted. This allows for a flexible trade-off between objective function optimization and queue backlog control.

[0144] Furthermore, this embodiment of the invention also provides a drone mission offloading device for a low-Earth orbit (LEO) satellite network. The LEO satellite network includes multiple LEO satellites, multiple drones, and a ground network. Each LEO satellite is interconnected via inter-satellite links and communicates with both the drones and the ground network. See also Figure 7 The unmanned aerial vehicle (UAV) mission offloading device 300 for this low-Earth orbit satellite network includes: Model building unit 310 is used to construct a joint optimization model based on the low-Earth orbit satellite network with the objective of maximizing the long-term average task offload. The joint optimization model includes the long-term average computing energy consumption constraint of the low-Earth orbit satellite and the long-term average cost constraint of the low-Earth orbit satellite network to wholesale computing resources to the ground network.

[0145] The queue construction unit 320 is used to construct a virtual computing energy consumption queue and a virtual resource wholesale cost queue for the long-term average computing energy consumption constraint and the long-term average cost constraint, respectively. The virtual computing energy consumption queue represents the cumulative deviation of the actual computing energy consumption of each low-Earth orbit satellite in each time slot from its long-term average energy consumption constraint, and the virtual resource wholesale cost queue represents the cumulative deviation of the actual resource wholesale cost of the low-Earth orbit satellite network in each time slot from its long-term average cost constraint.

[0146] Model transformation unit 330 is used to transform the joint optimization model into a single-slot static optimization problem based on the Lyapunov optimization framework. The transformation is achieved by minimizing the single-slot drift plus penalty function, which includes the task unloading benefit term of the current slot, the penalty term for virtual queue backlog, and Lyapunov weight coefficients used to adjust the tradeoff between task processing performance and the degree of long-term constraint satisfaction.

[0147] The problem-solving unit 340 is used to solve the single-time slot static optimization problem in each time slot based on the arrival status of UAV tasks and the backlog status of the virtual queue in the current time slot, and obtain the service deployment decision, computing resource allocation decision and task offloading decision to the ground network for that time slot.

[0148] In summary, the embodiments of this invention provide a method, apparatus, electronic device, and storage medium for unmanned aerial vehicle (UAV) mission offloading in a low-Earth orbit (LEO) satellite network. By constructing a dual virtual queue to dynamically represent long-term average computational energy consumption constraints and long-term average resource wholesale cost constraints, and based on the Lyapunov optimization framework, transforming the long-term objective coupled across time slots into a single-time slot static optimization problem, it achieves integrated online solution for service deployment, computational resource allocation, and satellite-to-ground mission offloading decisions under a three-level collaborative architecture of UAVs, LEO satellites, and ground networks without needing to predict future mission arrival distribution. Furthermore, it generates relaxed service deployment decisions through deep neural networks and integrates greedy initialization and K-nearest neighbor enhancement mechanisms to generate feasible candidate sets that satisfy the hard constraints of satellite storage capacity, overcoming the real-time bottleneck caused by strong coupling between discrete and continuous variables in mixed-integer nonlinear programming.

[0149] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0150] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0151] If the functionality is implemented as a software module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a computer-readable storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0152] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0153] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for UAV task offloading of a low earth orbit satellite network, the method comprising: The low-Earth orbit (LEO) satellite network includes multiple LEO satellites, multiple UAVs, and a ground network. Each LEO satellite is interconnected via inter-satellite links and communicates with both the UAVs and the ground network. The method includes: Based on the aforementioned low-Earth orbit satellite network, a joint optimization model is constructed with the objective of maximizing the long-term average task offload. The joint optimization model includes the long-term average computing energy consumption constraint of the low-Earth orbit satellites and the long-term average cost constraint of the low-Earth orbit satellite network wholesaling computing resources to the ground network. A virtual computing energy consumption queue and a virtual resource wholesale cost queue are constructed for the long-term average computing energy consumption constraint and the long-term average cost constraint, respectively. Based on the Lyapunov optimization framework, the joint optimization model is transformed into a single-slot static optimization problem; Within each time slot, based on the arrival status of UAV missions and the backlog status of the virtual queue in the current time slot, the single-time slot static optimization problem is solved to obtain the service deployment decision, computing resource allocation decision, and mission offloading decision to the ground network for that time slot.

2. The method of claim 1, wherein, The construction of a virtual computing energy consumption queue and a virtual resource wholesale cost queue for the long-term average computing energy consumption constraint and the long-term average cost constraint, respectively, includes: The virtual computational energy consumption queue for each low-Earth orbit satellite is updated according to the following formula: in, Indicates low-orbit satellites In the time slot The backlog length of the virtual computing energy consumption queue. For low-orbit satellites The long-term average energy consumption threshold, For low-orbit satellites For drones Type Computing resources allocated to the task The energy consumption coefficient corresponding to the task performed by the satellite unit. This is the energy consumption factor amplification factor. The length of each time slot, For low-Earth orbit satellites, For time slots A collection of drones covered by an inner low-Earth orbit satellite network; Update the virtual resource wholesale cost queue for the entire low-Earth orbit satellite network according to the following formula: in, For time slots The backlog length of the virtual wholesale cost queue, The long-term average cost threshold. Unmanned aerial vehicle (UAV) terminals for offloading from low-Earth orbit (LEO) satellite networks to terrestrial networks The workload, For drone terminals In the time slot Chinese type The total arrival requirements of the task. Calculate the unit resource price for the terrestrial network. This is the wholesale resource cost ratio coefficient.

3. The method according to claim 2, characterized in that, The method based on the Lyapunov optimization framework transforms the joint optimization model into a single-slot static optimization problem, including: The Lyapunov function is defined as half the sum of the squares of all the virtual computing energy consumption queues and the virtual resource wholesale cost queues; The single-slot Lyapunov offset function is defined as the expected increment of the Lyapunov function of the current slot relative to the previous slot; The Lyapunov offset plus penalty function is constructed as the single-slot Lyapunov drift minus the product of the Lyapunov weight coefficient and the expected total task unloading amount in the current slot; By minimizing the upper bound of the Lyapunov offset plus penalty function, the joint optimization model is transformed into a single-slot static optimization problem that depends only on the current slot state and the virtual queue backlog in each slot.

4. The method according to claim 1, characterized in that, Within each time slot, based on the arrival status of UAV missions and the backlog status of the virtual queue in the current time slot, the single-time-slot static optimization problem is solved to obtain the service deployment decision, computing resource allocation decision, and mission offloading decision to the ground network for that time slot, including: The arrival status of drone missions and the backlog status of virtual queues in the current time slot are input into a pre-trained deep neural network to obtain relaxed service deployment decisions. The relaxed service deployment decision is extended to a set of candidate service deployment decisions that satisfy the satellite storage capacity constraint; For each of the candidate service deployment decisions, after fixing the service deployment decision, solve the subproblems involving only continuous variables in the single time slot static optimization problem to obtain the target reward value and the corresponding task unloading decision and computing resource allocation decision; The candidate service deployment decision with the largest target reward value, along with its corresponding task unloading decision and computing resource allocation decision, is selected as the final decision for the current time slot.

5. The method according to claim 4, characterized in that, The extension of the relaxed service deployment decision to a set of candidate service deployment decisions that satisfy satellite storage capacity constraints includes: The relaxed service deployment decision is divided into multiple segments according to a preset method to obtain multiple candidate action sub-vectors; wherein, the length of each candidate action sub-vector is the number of service types, and each candidate action sub-vector corresponds to the relaxation value of all service types on a low-Earth orbit satellite; For each candidate action subvector, an initial feasible service deployment decision is generated using a greedy strategy. An enhancement approach based on K-nearest neighbors is adopted to generate multiple new feasible service deployment decisions based on the initial feasible service deployment decision; All generated feasible service deployment decisions are used as candidate service deployment decisions.

6. The method according to claim 5, characterized in that, The K-nearest neighbor-based enhancement method generates multiple new feasible service deployment decisions based on the initial feasible service deployment decision, including: Based on the initial feasible service deployment decision, initialize the set of deployed services, the set of undeployed services, and the remaining capacity; Construct a first sorting queue and a second sorting queue; wherein, the first sorting queue is formed by calculating the absolute value of the relaxation value difference between each deployed service and each undeployed service, and sorting them in ascending order of the absolute value, to represent the nearest neighbor relationship between deployed services and undeployed services; the second sorting queue is formed by calculating the absolute value of the relaxation value difference between each undeployed service and each deployed service, and sorting them in ascending order of the absolute value, to represent the nearest neighbor relationship between undeployed services and deployed services. Set a preset number of neighbors K, select the first K deployed services from the first sorting queue, and select the first K undeployed services from the second sorting queue; For each selected deployed service and each undeployed service, exchange their deployment states and generate a new service deployment decision; Each time a new service deployment decision is generated, it is checked whether it meets the satellite storage capacity constraint. If it does, it is considered a new feasible service deployment decision.

7. The method according to claim 4, characterized in that, The method further includes: Store the current state obtained in each time slot and the selected optimal service deployment decision as a sample; Every preset period, a batch of samples is randomly selected, and the parameters of the deep neural network are updated using the Adam algorithm.

8. A drone mission unloading device for a low-Earth orbit satellite network, characterized in that, The low-Earth orbit (LEO) satellite network includes multiple LEO satellites, multiple UAVs, and a ground network. Each LEO satellite is interconnected via inter-satellite links and communicates with both the UAVs and the ground network. The device includes: The model building unit is used to construct a joint optimization model based on the low-Earth orbit satellite network with the objective of maximizing the long-term average task offload. The joint optimization model includes the long-term average computing energy consumption constraint of the low-Earth orbit satellite and the long-term average cost constraint of the low-Earth orbit satellite network to wholesale computing resources to the ground network. The queue construction unit is used to construct a virtual computing energy consumption queue and a virtual resource wholesale cost queue for the long-term average computing energy consumption constraint and the long-term average cost constraint, respectively. The virtual computing energy consumption queue represents the cumulative deviation of the actual computing energy consumption of each of the low-Earth orbit satellites from its long-term average energy consumption constraint in each time slot, and the virtual resource wholesale cost queue represents the cumulative deviation of the actual resource wholesale cost of the low-Earth orbit satellite network from its long-term average cost constraint in each time slot. The model transformation unit is used to transform the joint optimization model into a single-slot static optimization problem based on the Lyapunov optimization framework. The transformation is achieved by minimizing the single-slot drift plus penalty function, which includes the task unloading benefit term of the current slot, the penalty term for virtual queue backlog, and Lyapunov weight coefficients used to adjust the trade-off between task processing performance and the degree of long-term constraint satisfaction. The problem-solving unit is used to solve the single-time-slot static optimization problem in each time slot based on the arrival status of UAV tasks and the backlog status of the virtual queue in the current time slot, and obtain the service deployment decision, computing resource allocation decision and task offloading decision to the ground network for that time slot.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 7.

Images