Resource scheduling method and device, electronic equipment and storage medium

Abstract

The invention provides a resource scheduling method and device, electronic equipment and a storage medium, and the method comprises the steps that a resource scheduling optimization model takes minimization of system comprehensive cost formed by task queue backlog of a mobile terminal and resource consumption of the mobile terminal and an edge computing node as an optimization target; the availability of a mobile terminal, an edge computing node and communication resources is used as a constraint condition for construction; solving the resource scheduling optimization model to obtain a resource scheduling decision of the computing task of the mobile terminal; and related resources are controlled based on the resource scheduling decision. According to the method, the task delay and the long-term average value of the system power consumption are taken into a unified optimization target, and the real-time resource availability is taken as a hard constraint to ensure that the scheduling decision has global optimality and physical feasibility, so that the dynamic adaptive capacity of resource scheduling is remarkably improved on the basis of ensuring the performability of the resource scheduling decision, and the resource scheduling efficiency is improved. The long-term average power consumption of the system is effectively reduced, and the continuous operation duration of the mobile terminal is prolonged.

Description

Technical Field

[0001] This invention relates to the field of resource scheduling technology, and in particular to a resource scheduling method, apparatus, electronic device and storage medium. Background Technology

[0002] With the rapid development of the low-altitude economy, drones and intelligent vehicles are increasingly collaborating in scenarios such as agricultural plant protection, power line inspection, emergency rescue, urban logistics, and traffic monitoring. In these tasks, drones or ground vehicles typically act as the task initiator. When drones are used as the task initiator, such as for track inspection or bridge inspection, they need to collect multimodal perception data in real time, including high-definition visible images, infrared thermal imaging, and laser point clouds, and perform online analysis of equipment status based on artificial intelligence models. When intelligent vehicles are used as the task initiator, such as inspection vehicles or unmanned delivery vehicles, they are typically equipped with multi-source sensors such as cameras, lidar, and millimeter-wave radar to collect environmental data in real time.

[0003] However, limited by the power, energy consumption, and computing power of airborne and vehicle-mounted computing units, small inspection drones or intelligent vehicles struggle to efficiently run complex AI algorithms (such as Transformer and 3D point cloud segmentation) and maintain long-term operation. Uploading all raw data to a remote cloud center leads to high latency, high bandwidth consumption, and network congestion, especially in remote locations or densely populated urban areas with poor communication conditions. To address this challenge, Mobile Edge Computing (MEC) technology has been introduced into low-altitude operation systems. By deploying edge servers on ground base stations or high-altitude auxiliary platforms, low-latency, high-energy-efficiency computing services can be provided to inspection drones or vehicles locally. Recent research proposes using high-altitude drones as mobile MEC nodes, collaborating with ground macro base stations (providing stable and high-performance MEC services) to form an "integrated air-ground edge computing network," supporting dynamic offloading and parallel processing of inspection tasks.

[0004] Current computational offloading schemes for low-altitude inspection still have significant shortcomings. On the one hand, the offloading strategies are relatively static, lacking dynamic adaptability and stability. Existing systems mostly rely on preset rules (such as "upload if data volume exceeds a threshold") for task offloading, making it difficult to dynamically adjust decisions based on real-time changes in channel quality, device power consumption, queue load stability, and task priority, resulting in low resource utilization or task response timeouts. On the other hand, existing solutions lack a global energy efficiency optimization mechanism, typically only optimizing the power consumption of the UAV or base station individually, failing to fully consider the coupling relationship between communication power consumption and computational power consumption. This prevents the minimization of system-level power consumption, limiting the coverage and continuous operation time of a single inspection task. Summary of the Invention

[0005] This invention provides a resource scheduling method, apparatus, electronic device, and storage medium to address the shortcomings of current computational offloading schemes for low-altitude inspections, which are characterized by static offloading strategies, lack of dynamic adaptability and stability, and lack of a global energy efficiency optimization mechanism.

[0006] This invention provides a resource scheduling method, comprising the following steps: A resource scheduling optimization model is obtained. The optimization objective of the resource scheduling optimization model is to minimize the overall system cost consisting of the task queue backlog of the mobile terminal and the resource consumption of the mobile terminal and the edge computing node. The model is constructed with the availability of computing resources of the mobile terminal, the availability of computing resources of the edge computing node, and the availability of communication resources for the mobile terminal to offload tasks as constraints. Solve the resource scheduling optimization model to obtain the resource scheduling decision for the computing tasks of the mobile terminal; Based on the resource scheduling decision, the computing resources of the mobile terminal, the computing resources of the edge computing node, and the communication resources are controlled.

[0007] According to a resource scheduling method provided by the present invention, the availability of computing resources of the mobile terminal is determined based on the computing frequency of the mobile terminal; The availability of computing resources for the edge computing node is determined based on the computing frequency of the edge computing node; The availability of the communication resources is determined based on the proportion of the total communication bandwidth allocated to the mobile terminal.

[0008] The resource scheduling method provided by the present invention further includes delay constraints; The latency constraint is determined based on the average length of the task queues of the mobile terminal and the edge computing node, respectively.

[0009] According to a resource scheduling method provided by the present invention, the step of determining the overall system cost includes: Based on the task queue length of the mobile terminal and the task queue length of the edge computing node, the system queue backlog is determined; Based on the computing and communication power consumption of the mobile terminal and the computing power consumption of the edge computing node, the system resource consumption is determined. The overall cost of the system is determined based on the system queue backlog and the system resource consumption.

[0010] According to a resource scheduling method provided by the present invention, the step of solving the resource scheduling optimization model to obtain the resource scheduling decision for the computing task of the mobile terminal includes: Based on the resource scheduling optimization model, a single-slot optimization problem is constructed with the goal of minimizing the weighted sum of the changes in the system queue backlog and the system resource consumption. Obtain the system state information of the single-slot optimization problem; The system state information is input into a preset deep reinforcement learning network to obtain the resource scheduling decision output by the deep reinforcement learning network. The system status information includes current task queue information representing the backlog in the system queue, and channel status information representing the transmission quality of the communication resources.

[0011] According to a resource scheduling method provided by the present invention, the resource scheduling decision includes a task offloading ratio vector; The task offloading ratio vector is used to determine the proportion of data volume executed by the computing task in the mobile terminal, the airborne drone base station, and the ground macro base station.

[0012] According to a resource scheduling method provided by the present invention, the edge computing node includes a ground macro base station deployed in a ground cellular network and an aerial UAV base station carried by a UAV.

[0013] The present invention also provides a resource scheduling device, comprising the following units: A construction unit is used to obtain a resource scheduling optimization model. The resource scheduling optimization model takes minimizing the overall system cost consisting of the task queue backlog of the mobile terminal and the resource consumption of the mobile terminal and the edge computing node as the optimization objective, and is constructed with the availability of computing resources of the mobile terminal, the availability of computing resources of the edge computing node, and the availability of communication resources for the mobile terminal to offload tasks as constraints. The solution unit is used to solve the resource scheduling optimization model to obtain the resource scheduling decision for the computing tasks of the mobile terminal; The control unit is used to control the computing resources of the mobile terminal, the computing resources of the edge computing node, and the communication resources based on the resource scheduling decision.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the resource scheduling method described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the resource scheduling method as described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements any of the resource scheduling methods described above.

[0017] The resource scheduling method, apparatus, electronic device, and storage medium provided by this invention have a resource scheduling optimization model that aims to minimize the overall system cost comprised of task queue backlog on mobile terminals and resource consumption by mobile terminals and edge computing nodes. This model is constructed with the availability of mobile terminals, edge computing nodes, and communication resources as constraints. Solving the resource scheduling optimization model yields resource scheduling decisions for mobile terminal computing tasks. Based on these decisions, relevant resources are controlled. This method incorporates the long-term average values ​​of task latency and system power consumption into a unified optimization objective, while using real-time resource availability as a hard constraint, ensuring that scheduling decisions possess both global optimality and physical feasibility. This solves the problems of impractical decisions and system instability caused by static strategies and lack of consideration for resource constraints in existing technologies, as well as the inability to minimize total system power consumption due to the lack of global energy efficiency optimization. Therefore, while ensuring the executability of resource scheduling decisions, it significantly improves the dynamic adaptability of resource scheduling, effectively reduces the long-term average system power consumption, and extends the continuous operation time of mobile terminals. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the resource scheduling method provided by the present invention.

[0020] Figure 2 This is a schematic diagram of the resource scheduling system provided by the present invention.

[0021] Figure 3 This is a flowchart illustrating the resource scheduling algorithm based on Lyapunov optimization and deep reinforcement learning provided by this invention.

[0022] Figure 4 This is a schematic diagram of the resource scheduling device provided by the present invention.

[0023] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0025] This invention provides a resource scheduling method. The core idea of ​​this invention is to construct and solve a resource scheduling optimization model with the goal of minimizing the overall system cost, thereby obtaining the optimal resource scheduling decision, and intelligently controlling mobile terminals, edge computing nodes, and communication resources based on this decision. This method can effectively balance task processing latency and system energy consumption, improving overall operating efficiency and stability in complex dynamic environments (such as low-altitude intelligent inspection scenarios). This method can be applied to mobile edge computing (MEC) systems, especially suitable for air-ground collaborative computing systems that include smart mobile devices (SMDs), which may include mobile terminals and edge computing nodes. The mobile edge computing system may include intelligent vehicles or inspection drones as mobile terminals, and ground-based macro base stations (MBS) or mobile base stations (mBS), such as aerial drone base stations, as edge computing nodes. When performing computationally intensive tasks such as high-definition image analysis and 3D point cloud segmentation, mobile terminals can offload some or all of the task data to edge computing nodes for processing to compensate for their own limitations in computing power, power consumption, or energy consumption. The method of this invention aims to efficiently and intelligently schedule computing and communication resources in this process. Figure 1 This is a flowchart illustrating the resource scheduling method provided by the present invention. Figure 2 This is a schematic diagram of the resource scheduling system provided by the present invention, such as... Figure 1 , Figure 2 As shown, the method includes steps 110, 120 and 130.

[0026] Step 110: Obtain the resource scheduling optimization model. The resource scheduling optimization model is constructed with the goal of minimizing the overall system cost consisting of the task queue backlog of the mobile terminal and the resource consumption of the mobile terminal and the edge computing node, and with the availability of computing resources of the mobile terminal, the availability of computing resources of the edge computing node, and the availability of communication resources for the mobile terminal to offload tasks as constraints.

[0027] Specifically, firstly, a resource scheduling optimization model can be obtained, and the method for obtaining this model can be varied. For example, in a centralized control system, a central controller, such as one deployed on the core network or a main base station, can construct and maintain the resource scheduling optimization model based on a preset system architecture and parameters. In a distributed system, individual nodes, such as edge computing nodes, can also collaboratively construct the resource scheduling optimization model based on collected global or local information. Alternatively, the resource scheduling optimization model can be embedded as a pre-built algorithm module in the system controller and invoked or loaded when the system starts or at the beginning of a specific scheduling cycle. This embodiment of the invention does not specifically limit this approach.

[0028] The resource scheduling optimization model aims to minimize the overall system cost, which consists of the backlog of task queues on mobile terminals and the resource consumption of mobile terminals and edge computing nodes. It is constructed with constraints on the availability of computing resources on mobile terminals, the availability of computing resources on edge computing nodes, and the availability of communication resources for task offloading on mobile terminals.

[0029] Here, the task queue backlog on the mobile terminal reflects the system's service quality and the immediacy of task processing. Task queue backlog can be understood as the accumulation of the number of tasks or the amount of data waiting to be processed or unloaded locally on the mobile terminal. It's understandable that a larger task queue backlog on a mobile terminal means a longer average waiting time for tasks, greater latency, a worse user experience, and may even lead to task timeouts or system instability. The task queue backlog on the mobile terminal can include the task queue locally on the mobile terminal, and can further include task queues that have been transferred to edge computing nodes for processing due to unloading.

[0030] Here, the resource consumption of mobile terminals and edge computing nodes reflects the system's operating costs, especially energy consumption. For battery-powered mobile terminals, such as drones and smart cars, reducing resource consumption directly affects their battery life and operating time. Resource consumption can specifically include the computing power consumed by the mobile terminal performing local computation, the communication power consumed by the mobile terminal for task offloading, and the computing power consumed by the edge computing nodes for processing offloading tasks. Therefore, minimizing resource consumption is key to achieving green and long-term system operation.

[0031] The availability of computing resources for a mobile terminal refers to the maximum computing power that the mobile terminal can provide at any given moment. This availability can be determined by factors such as the maximum computing frequency of the mobile terminal's processor, its current workload, and battery level. The amount of local computing tasks allocated to the mobile terminal by the scheduling decision cannot exceed its capabilities.

[0032] The availability of computing resources for edge computing nodes refers to the computing power that edge computing nodes, such as terrestrial macro base stations or drone base stations, can provide. Here, the availability of computing resources for edge computing nodes is typically determined by the configuration of the edge computing node's servers and the current workload. The total number of offloaded tasks allocated to an edge computing node cannot exceed its processing capacity limit.

[0033] The availability of communication resources for task offloading on mobile terminals refers to the quality and capacity of the communication link between the mobile terminal and the edge computing node. Here, the availability of communication resources can be determined by factors such as the channel gain, available bandwidth, and signal-to-noise ratio of the wireless channel; however, this embodiment of the invention does not impose specific limitations on these factors. The availability of communication resources directly limits the amount of data that can be successfully offloaded per unit time.

[0034] Step 120: Solve the resource scheduling optimization model to obtain the resource scheduling decision for the computing tasks of the mobile terminal.

[0035] Specifically, after obtaining the resource scheduling optimization model, the resource scheduling optimization model can be solved to obtain the resource scheduling decision for the computing tasks of the mobile terminal.

[0036] Here, the methods for solving the resource scheduling optimization model can be diverse, and their selection depends on the specific mathematical form of the resource scheduling optimization model. For example, for complex nonlinear resource scheduling optimization models that include random variables, advanced solution algorithms such as convex optimization, heuristic algorithms, or machine learning-based algorithms can be used. In a preferred embodiment, the resource scheduling optimization model can use an algorithm that combines Lyapunov optimization with deep reinforcement learning.

[0037] Step 130: Based on the resource scheduling decision, control the computing resources of the mobile terminal, the computing resources of the edge computing node, and the communication resources.

[0038] Specifically, after obtaining the resource scheduling decision, the computing resources of the mobile terminal, the computing resources of the edge computing node, and the resources used for communication can be controlled based on the resource scheduling decision.

[0039] This includes controlling the computing resources of the mobile terminal. For example, the controller sends instructions to the mobile terminal to adjust the operating frequency of the mobile terminal's CPU (Central Processing Unit) to match the local computing load specified in the decision, thereby saving energy while meeting computing needs.

[0040] The controller controls the computing resources of edge computing nodes. For example, the controller reserves and allocates the necessary virtual machine or container resources for the unloading task from the mobile terminal on the corresponding edge computing node, and sets the corresponding computing capacity.

[0041] Communication resources are controlled, for example, the controller allocates specified wireless channels and bandwidth resources to the mobile terminal's offloading process based on the decision, and the mobile terminal adjusts its transmission power based on the decision to ensure that the mission data can be successfully transmitted to the target edge computing node at the rate and quantity planned in the decision.

[0042] By implementing the above steps, the resource scheduling method of the present invention can complete a complete closed loop within one scheduling cycle, from sensing the system state, establishing the optimization problem, solving the optimal decision, to executing resource control.

[0043] The method provided in this invention uses a resource scheduling optimization model with the optimization objective of minimizing the overall system cost comprised of task queue backlog on mobile terminals and resource consumption by mobile terminals and edge computing nodes, and constructs a model with the availability of mobile terminals, edge computing nodes, and communication resources as constraints. Solving the resource scheduling optimization model yields resource scheduling decisions for mobile terminal computing tasks. Based on these resource scheduling decisions, relevant resources are controlled. This method incorporates the long-term average values ​​of task latency and system power consumption into a unified optimization objective, while using real-time resource availability as a hard constraint, ensuring that scheduling decisions possess both global optimality and physical feasibility. This solves the problems of existing technologies, such as impractical decisions and system instability due to static strategies and lack of consideration for resource constraints, and the inability to minimize total system power consumption due to the lack of global energy efficiency optimization. Therefore, while ensuring the executability of resource scheduling decisions, it significantly improves the dynamic adaptability of resource scheduling, effectively reduces the long-term average system power consumption, and extends the continuous operation time of mobile terminals.

[0044] Based on the above embodiments, the availability of computing resources of the mobile terminal is determined based on the computing frequency of the mobile terminal; The availability of computing resources for the edge computing node is determined based on the computing frequency of the edge computing node; The availability of the communication resources is determined based on the proportion of the total communication bandwidth allocated to the mobile terminal.

[0045] Specifically, the availability of computing resources for a mobile terminal is determined based on its computing frequency. The computing frequency of a mobile terminal refers to the operating frequency of its internal processor, which can be measured in Hertz (Hz). The computing frequency directly determines the number of computational instructions the mobile terminal can execute per unit of time, thus characterizing its computing power. For example, a mobile terminal's maximum computing frequency might be F... localmaxUnder any scheduling decision, the frequency f allocated to it for local computation local None of them can exceed this upper limit, that is, 0 ≤ f local ≤ F localmax By directly linking the availability of computing resources to the physically controllable parameter of computing frequency, precise control over computing resources can be achieved.

[0046] Similarly, the availability of computing resources for edge computing nodes is determined based on their computing frequency. Each edge computing node also has one or more processors. Its computing resource availability can also be characterized by the computing frequency it can provide. For example, for a mobile terminal... i Edge servers allocated computing frequencies, their allocated computing frequencies It cannot exceed the total available computing frequency of the edge servers, such as .

[0047] Here, the availability of communication resources is determined based on the proportion of total communication bandwidth allocated to mobile terminals. In wireless communication systems, the total available bandwidth resources are finite. The total communication bandwidth proportion refers to the percentage of bandwidth resources allocated from the total available bandwidth within a scheduling cycle, specifically for a particular mobile terminal to perform task data upload (unload). For example, in a system using Orthogonal Frequency Division Multiple Access (OFDMA), the total bandwidth consists of N sub-channels. If k sub-channels are allocated to a mobile terminal, then its bandwidth proportion is k / N.

[0048] The proportion of total communication bandwidth directly affects the data transmission rate of mobile terminals. Defining the availability of communication resources as bandwidth allows scheduling decisions to be translated into specific instructions for the base station resource allocation unit, i.e., allocating a specific proportion of spectrum resources to a specific user.

[0049] Based on the above embodiments, delay constraints are also included; The latency constraint is determined based on the average length of the task queues of the mobile terminal and the edge computing node, respectively.

[0050] Specifically, latency constraints are determined based on the average length of the task queues of both the mobile terminal and the edge computing node. These latency constraints aim to balance system energy efficiency and service quality, ensuring that while pursuing maximum energy efficiency, processing latency is prevented from increasing indefinitely due to queue congestion, thereby guaranteeing basic service timeliness.

[0051] According to Little's rule in queuing theory, in a stable system, the average number of tasks in a queue (i.e., the average queue length) is proportional to the average waiting time of the tasks (i.e., the average delay). Therefore, by imposing an upper limit on the average length of the task queue, an upper limit can be indirectly imposed on the average delay of the tasks.

[0052] Therefore, a specific way to implement delay constraints is to set a length threshold for each task queue, ensuring that the instantaneous length of the queue is within this threshold at any given time. For example, for mobile terminals... i Local task queue Its task queue on the airborne drone base station and its task queue on ground macro base stations. All must meet The constraints, among which This is a preset queue length threshold to ensure user experience.

[0053] Furthermore, to ensure the long-term stable operation of the system and avoid the infinite accumulation of tasks, the resource scheduling optimization model of this invention can also introduce queue stability constraints. The queue stability constraints require that the long-term time average of all task queues be bounded.

[0054] In a preferred embodiment, it is required to converge to a finite value, ideally approaching zero. (Using a mobile terminal) i Taking the local queue as an example, this constraint can be mathematically expressed as: Let T be a bounded constant, where T is the total number of time slots. This represents the expected operation. This constraint also applies to all task queues deployed on edge computing nodes.

[0055] Based on the above embodiments, the step of determining the overall system cost includes: Step 210: Determine the system queue backlog based on the task queue length of the mobile terminal and the task queue length of the edge computing node; Step 220: Determine system resource consumption based on the computing power consumption and communication power consumption of the mobile terminal and the computing power consumption of the edge computing node; Step 230: Determine the overall system cost based on the system queue backlog and the system resource consumption.

[0056] Specifically, in a more concrete mathematical model, the dynamic evolution of the system can be modeled in discrete time slots, each with a length of ς. It can be assumed that in each scheduling time slot t, each mobile terminal... i Each of these will independently generate a new computational task, with a data volume of [data missing]. The arrival process of a task can be modeled as a stochastic process that follows a specific probability distribution; for example, it can be assumed that... It follows a Poisson distribution.

[0057] First, the system queue backlog can be determined based on the task queue lengths of the mobile terminal and the edge computing nodes. System queue backlog is an indicator that measures the accumulation of tasks awaiting processing throughout the system. At a specific time t, it can be quantified as the sum of the lengths of all relevant queues. For example, if the system includes a mobile terminal, an aerial drone base station, and a ground macro base station, then the system queue backlog Qsys(t) can be defined as: Qsys(t) = Qlocal(t) + Quav(t) + Qmbs(t), where Qlocal(t) is the length of the mobile terminal's local task queue, and Quav(t) and Qmbs(t) are the lengths of the task queues originating from the mobile terminal that are waiting to be processed at the aerial drone base station and the ground macro base station, respectively.

[0058] Specifically, each computing unit (locally on the mobile terminal and at each edge node) maintains an independent task queue for tasks originating from a specific mobile terminal. For example, the mobile terminal... i The local task queue is denoted as In time slot t, if local computing power is insufficient to handle all pending tasks, the unfinished portions will be cached. Meanwhile, tasks offloaded to the aerial drone base station will enter a dedicated queue thereon. The lengths of these queues will be dynamically updated in each time slot based on task arrival, processing, and unloading. For example, queues... Updates can follow these rules: in, It is in time slot t from the mobile terminal i The amount of task data uploaded to mBS, and The mBS is for the mobile terminal in this time slot. i The actual amount of data processed. Similarly, local queues. The queues on other edge nodes also follow similar update logic.

[0059] Secondly, based on the computing and communication power consumption of the mobile terminal and the computing power consumption of the edge computing nodes, the system resource consumption is determined. System resource consumption refers to the total power consumption of the system. It can be quantified as the sum of the power consumption of each component.

[0060] In a specific physical model, the various costs can be quantified as follows: The computing power consumption of a mobile terminal is primarily determined by the computing frequency of its processor. According to circuit theory, local computing power consumption... It can be derived from the formula Confirmed, among which It is the computing frequency allocated for local computing. This is an effective switched-capacitor coefficient that depends on the mobile terminal chip's circuit architecture. Simultaneously, it represents the amount of task data that can be processed locally within this time slot. It can be represented as ,in It deals with density.

[0061] The power consumption of edge computing nodes is similar to that of local computing. Edge nodes (such as mBS) process data from mobile terminals. i The computational power consumed by the task It can also be modeled as ,in Is the edge node assigned to the task? i The calculation frequency, This is the effective switching capacitor coefficient of its server. Its workload... Confirmed, among which This represents the processing density of the edge server. The computational power consumption and processing model of the terrestrial macro base station (MBS) can also be established in the same way.

[0062] The communication power consumption of a mobile terminal mainly occurs during task offloading. Its magnitude can be determined based on the classic Shannon-Hartley theorem, and it is positively correlated with the mobile terminal's transmit power. To achieve a certain transmission rate, the required transmit power depends on the allocated channel bandwidth and the channel quality between the mobile terminal and the target edge node, the latter being affected by the channel gain. The impact of factors such as path loss.

[0063] The computing power consumption of a mobile terminal is typically proportional to a high power of the computing frequency allocated to it. The communication power consumption of a mobile terminal mainly refers to the transmission power consumption incurred for offloading data, which depends on the allocated transmission power. The computing power consumption of an edge computing node is proportional to the computing frequency allocated to offloading tasks.

[0064] Finally, the overall system cost is determined based on system queue backlog and system resource consumption. The most common method for determining this cost is by weighted summation. For example, the overall system cost C(t) within a scheduling slot t can be expressed as: C(t) = Qsys(t) + V * Psys(t), where V is a non-negative weighting coefficient.

[0065] Based on the above embodiments, step 120 includes: Step 121: Based on the resource scheduling optimization model, construct a single-slot optimization problem with the objective of minimizing the weighted sum of the changes in the system queue backlog and the system resource consumption; Step 122: Obtain the system state information of the single-slot optimization problem; Step 123: Input the system status information into a preset deep reinforcement learning network to obtain the resource scheduling decision output by the deep reinforcement learning network; The system status information includes current task queue information representing the backlog in the system queue, and channel status information representing the transmission quality of the communication resources.

[0066] Specifically, the first step can be to construct a single-slot optimization problem based on a resource scheduling optimization model, with the objective of minimizing the weighted sum of the changes in system queue backlog and system resource consumption. This construction method is a typical application of Lyapunov optimization theory, and its core idea is to decompose a long-term, time-coupled average cost minimization problem into a series of independent, time-slot-wise deterministic optimization problems.

[0067] The objective function of this single-slot problem is typically expressed as a drift penalty term, in the form Δ(Qsys(t)) + V×Psys(t). Here, Δ(Qsys(t)) represents the conditional drift of the Lyapunov function defined based on the system queue backlog Qsys(t) within time slot t, used to measure queue stability; V×Psys(t) is a weighted term for the system resource consumption cost within time slot t, with the parameter V controlling the trade-off between cost minimization and queue stability. By greedily minimizing this drift penalty term in each time slot, the strong stability of all queues in the system can be theoretically guaranteed, and the long-term average system cost converges to a bounded interval.

[0068] The second step is to obtain the system state information for the single-slot optimization problem. To make optimal decisions, the deep reinforcement learning network needs to fully understand the current system state. This system state information can be input into a pre-defined deep reinforcement learning network to obtain the resource scheduling decision output by the network.

[0069] The system state information includes current task queue information, which characterizes the system's queue backlog, and channel state information, which characterizes the transmission quality of communication resources. The current task queue information can be a vector containing the local queue length of the mobile terminal and the relevant task queue lengths on each edge computing node, for example, Squeue(t) = [Qlocal(t), Quav(t), Qmbs(t)]. The current task queue information reflects the degree of system congestion and is the basis for calculating the drift term.

[0070] Channel state information can be a vector of the wireless channel gain between the mobile terminal and each edge computing node. Channel quality directly determines the data offloading rate and power consumption, and is a key basis for communication resource allocation and offloading decisions.

[0071] The preset deep reinforcement learning network is a trained neural network model, such as a network constructed by algorithms such as deep Q-network, deep deterministic policy gradient, or proximal policy optimization. This embodiment of the invention does not specifically limit this network.

[0072] The deep reinforcement learning network receives the aforementioned system state information as input and, through its internal multi-layer nonlinear transformations, outputs a resource scheduling decision that minimizes the single-slot optimization objective (drift penalty term). This resource scheduling decision is an action vector containing multiple control parameters, such as determining the task offloading ratio, the allocation of computing frequencies to each endpoint, and the allocation of communication bandwidth and power. The decision-making process of the deep reinforcement learning network can be designed to be phased or layered. For example, the deep reinforcement learning network can first make a first-level decision based on channel state information and current task queue information, namely, making a crucial channel allocation decision, such as deciding whether to primarily use the channel with the UAV base station or the channel with the ground station. After determining this basic decision, subsequent parts of the network or another network module then make a second-level decision, that is, under this channel allocation scheme, jointly optimizing and solving for the remaining decision variables, such as the CPU computing frequency of each endpoint, the specific amount of offloaded data, and the fine-grained bandwidth allocation ratio, thus ultimately forming a complete resource scheduling decision.

[0073] The method provided in this invention, by combining Lyapunov optimization with deep reinforcement learning, can effectively solve resource scheduling problems in dynamic and complex environments. Lyapunov theory transforms long-term problems into short-sighted, single-slot problems, significantly reducing solution complexity; while deep reinforcement learning, with its powerful function approximation and learning capabilities, can learn near-optimal scheduling strategies from high-dimensional state information without requiring an exact system model. This method is not only computationally efficient, meeting the needs of real-time decision-making, but also highly adaptable to dynamic environmental changes.

[0074] Based on the above embodiments, the resource scheduling decision includes a task offloading ratio vector; The task offloading ratio vector is used to determine the proportion of data volume executed by the computing task in the mobile terminal, the airborne drone base station, and the ground macro base station.

[0075] Specifically, resource scheduling decisions include a task offloading ratio vector. This task offloading ratio vector, denoted as a(t), is a vector composed of multiple components. Each component represents how newly arriving computational task data should be segmented and allocated to different computational units for processing within a scheduling time slot t.

[0076] The system comprises three available computing locations: mobile terminals, aerial drone base stations, and ground-based macro base stations. Therefore, the task offloading ratio vector can be a three-dimensional vector a(t) = [as(t), au(t), am(t)], where as(t) represents the proportion of task data processed locally on the mobile terminal relative to the total task data; au(t) represents the proportion of task data offloaded to the aerial drone base station relative to the total task data; and am(t) represents the proportion of task data offloaded to the ground-based macro base station relative to the total task data. All three ratio components are non-negative, and their sum must be 1, i.e., as(t) + au(t) + am(t) = 1.

[0077] When a new computational task of size L bits arrives, the system divides the task according to the task offloading ratio vector: as(t) * L bits of data are placed into the local processing queue of the mobile terminal; au(t) * L bits of data are packaged and prepared to be offloaded to the airborne drone base station via the wireless link; and am(t) * L bits of data are packaged and prepared to be offloaded to the ground macro base station.

[0078] The method provided in this invention achieves refined and divisible scheduling of computational tasks by introducing a task offloading ratio vector. This allows the system to move beyond the binary choice of local computation or complete offloading, and instead flexibly distribute a task among multiple computing units for parallel processing based on the current load, channel conditions, and energy consumption of each computing unit. This refined load balancing strategy can more fully utilize all available computing and communication resources in the system, achieving better system performance, such as finding a task partitioning and offloading scheme with the lowest energy consumption while satisfying latency constraints.

[0079] Based on the above embodiments, the edge computing node includes a ground macro base station deployed in a ground cellular network and an aerial drone base station carried by a drone.

[0080] Specifically, an edge computing node is a heterogeneous, air-ground collaborative combination, including ground macro base stations deployed in terrestrial cellular networks and aerial drone base stations carried by drones.

[0081] Among them, the terrestrial macro base station deployed in the terrestrial cellular network is a traditional cellular network infrastructure. In this architecture, in addition to undertaking the function of wide-area wireless coverage, the terrestrial macro base station also integrates or deploys powerful computing servers, namely MEC servers, in the near field.

[0082] Aerial drone base stations, carried by drones, represent a new and flexible aerial network platform. They integrate small base station equipment and a certain computing processing unit onto a single drone. Their advantages include: high mobility and flexibility, allowing them to dynamically adjust their position based on the location of mobile terminals, flying nearby to establish high-quality line-of-sight or near-line-of-sight communication links, thus achieving low-latency, low-power, and rapid data offloading; and rapid deployment to areas inaccessible by conventional base stations, such as remote mountainous regions and post-disaster emergency scenarios.

[0083] By combining ground-based macro base stations with aerial drone base stations, a complementary air-ground collaborative edge computing architecture has been constructed. In this heterogeneous system, the ground-based macro base station serves as the computing core, providing stable and powerful basic computing power; while the aerial drone base station acts as a flexible front-end node, enabling efficient near-field access and computing services.

[0084] The resource scheduling method proposed in this invention, based on the characteristics of this heterogeneous system, dynamically and intelligently determines the execution location of computing tasks according to the real-time task requirements of the mobile terminal, its geographical location, and the channel status between it and two types of base stations. Specifically, this method can decide whether to keep the task processed locally on the terminal, or offload it to a UAV base station with a closer access distance and faster response, or to a ground macro base station with stronger computing power but a longer transmission distance. It also supports the collaborative scheduling of resources from all three types after task segmentation, thereby achieving global efficiency optimization.

[0085] Based on any of the above embodiments, this invention proposes a resource scheduling method under an air-ground collaborative computing architecture. This architecture consists of intelligent vehicles or inspection drones as mobile terminals, aerial drone base stations (mBS) and ground macro base stations (MBS) as edge computing nodes, and a remote cloud center (CDC). Its core lies in minimizing the long-term average total power consumption of the system while ensuring task queue stability and meeting end-to-end latency constraints. This invention designs a dynamic task offloading decision mechanism based on real-time state awareness. This mechanism supports splitting a single task and offloading it in parallel to multiple computing nodes. Through an online resource scheduling algorithm combining Lyapunov optimization and deep reinforcement learning, it solves a joint optimization problem involving computation, communication, and transmission power consumption in real time, thereby dynamically determining the offloading ratio of tasks among the local machine, edge nodes, and the cloud center, and jointly allocating bandwidth and computing frequency to achieve load balancing. Compared with existing technologies, this invention can significantly reduce the total power consumption of the system and extend the battery life of mobile devices through intelligent scheduling. At the same time, by leveraging the flexible deployment of UAV base stations, it effectively enhances the robustness and coverage of the system, ensuring the continuity of computing services in complex terrain or signal blind spots, and ultimately achieving the unity of stable system operation and intelligent scheduling.

[0086] Figure 3This is a flowchart illustrating the resource scheduling algorithm based on Lyapunov optimization and deep reinforcement learning provided by this invention, as shown below. Figure 3 As shown, this algorithm aims to achieve real-time resource allocation with low complexity, high convergence, and high stability. At the beginning of each scheduling slot, the algorithm first initializes and obtains the current system state, which mainly includes the task queue state of each node, such as the queue lengths on the local machine, mBS, and MBS, as well as the wireless channel state between the mobile terminal and each base station. Based on Lyapunov optimization theory, this state is used to construct a single-slot deterministic subproblem with the objective of minimizing the Lyapunov drift plus power penalty term. In each scheduling slot, the channel gain, SMDs, mBS, and MBS task queue and virtual queue information are first input as the current system state. This state information is fed into the deep neural network (DNN) of the actor module, which outputs a continuous channel allocation action decision. To facilitate the search, the continuous action is processed by the actor module: the action quantizer, which generates a set of discrete candidate channel allocation decisions. arrive For each candidate channel allocation decision, the system processes them in parallel. The optimal solutions for the remaining decision variables, including bandwidth allocation, offload, and CPU frequency, are calculated. Then, through an ArgMin step, the system compares all parallel computation results to find the scheme that minimizes the single-slot optimization objective (i.e., Lyapunov drift plus power consumption) and selects the optimal action. This is the optimal action. Together with the other previously calculated optimal decision variables, they constitute the optimal decision variable combination for the system. , The system then makes decisions regarding channel allocation, bandwidth allocation, offload capacity, and CPU frequency allocation. After these decisions are executed, the queuing system processes data and consumes power. Simultaneously, the task data arrives at time slot t. The system updates the queuing system queue to form the initial state for the next time slot. Simultaneously, in the policy update module, the selected optimal channel decision and the current system observation state are recorded and stored in the replay memory pool. The training process then uses random batches of samples selected from this memory pool to continuously train the deep neural network (DNN), thereby constantly optimizing the decision-making ability of the actor module. Finally, the system determines whether the termination condition is met. If not, it returns to the starting point of the process to continue scheduling the next time slot; if yes, the process ends.

[0087] The resource scheduling device provided by the present invention is described below. The resource scheduling device described below and the resource scheduling method described above can be referred to in correspondence.

[0088] Based on any of the above embodiments Figure 4 This is a schematic diagram of the resource scheduling device provided by the present invention, as shown below. Figure 4 As shown, the device includes: The construction unit 410 is used to obtain a resource scheduling optimization model. The resource scheduling optimization model takes minimizing the overall system cost consisting of the task queue backlog of the mobile terminal and the resource consumption of the mobile terminal and the edge computing node as the optimization objective, and is constructed with the availability of computing resources of the mobile terminal, the availability of computing resources of the edge computing node, and the availability of communication resources for the mobile terminal to offload tasks as constraints. Solving unit 420 is used to solve the resource scheduling optimization model to obtain the resource scheduling decision for the computing tasks of the mobile terminal; The control unit 430 is used to control the computing resources of the mobile terminal, the computing resources of the edge computing node, and the communication resources based on the resource scheduling decision.

[0089] The apparatus provided in this invention uses a resource scheduling optimization model with the optimization objective of minimizing the overall system cost comprised of task queue backlog on mobile terminals and resource consumption by mobile terminals and edge computing nodes, and is constructed with the availability of mobile terminals, edge computing nodes, and communication resources as constraints. Solving the resource scheduling optimization model yields resource scheduling decisions for mobile terminal computing tasks; based on these decisions, relevant resources are controlled. This method incorporates the long-term average values ​​of task latency and system power consumption into a unified optimization objective, while using real-time resource availability as a hard constraint, ensuring that scheduling decisions possess both global optimality and physical feasibility. This solves the problems of impractical decisions and system instability caused by static strategies and lack of consideration for resource constraints in existing technologies, as well as the inability to minimize total system power consumption due to the lack of global energy efficiency optimization. Therefore, while ensuring the executability of resource scheduling decisions, it significantly improves the dynamic adaptability of resource scheduling, effectively reduces the long-term average system power consumption, and extends the continuous operation time of mobile terminals.

[0090] Based on any of the above embodiments, the availability of computing resources of the mobile terminal is determined based on the computing frequency of the mobile terminal; The availability of computing resources for the edge computing node is determined based on the computing frequency of the edge computing node; The availability of the communication resources is determined based on the proportion of the total communication bandwidth allocated to the mobile terminal.

[0091] Based on any of the above embodiments, a delay constraint is also included; The latency constraint is determined based on the average length of the task queues of the mobile terminal and the edge computing node, respectively.

[0092] Based on any of the above embodiments, a determining unit is further included, wherein the determining unit specifically includes: The first determining unit is used to determine the system queue backlog based on the task queue length of the mobile terminal and the task queue length of the edge computing node; The second determining unit is used to determine the system resource consumption based on the computing power consumption and communication power consumption of the mobile terminal and the computing power consumption of the edge computing node. The third determining unit is used to determine the overall cost of the system based on the system queue backlog and the system resource consumption.

[0093] Based on any of the above embodiments, the solving unit 420 is specifically used for: Based on the resource scheduling optimization model, a single-slot optimization problem is constructed with the goal of minimizing the weighted sum of the changes in the system queue backlog and the system resource consumption. Obtain the system state information of the single-slot optimization problem; The system state information is input into a preset deep reinforcement learning network to obtain the resource scheduling decision output by the deep reinforcement learning network. The system status information includes current task queue information representing the backlog in the system queue, and channel status information representing the transmission quality of the communication resources.

[0094] Based on any of the above embodiments, the resource scheduling decision includes a task offloading ratio vector; The task offloading ratio vector is used to determine the proportion of data volume executed by the computing task in the mobile terminal, the airborne drone base station, and the ground macro base station.

[0095] Based on any of the above embodiments, the edge computing node includes a ground macro base station deployed in a ground cellular network and an aerial drone base station carried by a drone.

[0096] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 5As shown, the electronic device may include a processor 510, a communications interface 520, a memory 530, and a communication bus 540, wherein the processor 510, communications interface 520, and memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a resource scheduling method. This method includes: obtaining a resource scheduling optimization model, the resource scheduling optimization model being constructed with the optimization objective of minimizing the overall system cost constituted by the task queue backlog of the mobile terminal and the resource consumption of the mobile terminal and edge computing nodes, and constrained by the availability of computing resources of the mobile terminal, the availability of computing resources of the edge computing nodes, and the availability of communication resources for task offloading of the mobile terminal; solving the resource scheduling optimization model to obtain resource scheduling decisions for the computing tasks of the mobile terminal; and controlling the computing resources of the mobile terminal, the computing resources of the edge computing nodes, and the communication resources based on the resource scheduling decisions.

[0097] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present 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 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 described in the various embodiments of the present 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.

[0098] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the resource scheduling method provided by the above methods. The method includes: obtaining a resource scheduling optimization model, wherein the resource scheduling optimization model is constructed with the goal of minimizing the overall system cost consisting of the task queue backlog of the mobile terminal and the resource consumption of the mobile terminal and the edge computing node, and with the availability of computing resources of the mobile terminal, the availability of computing resources of the edge computing node, and the availability of communication resources for the mobile terminal to offload tasks as constraints; solving the resource scheduling optimization model to obtain a resource scheduling decision for the computing tasks of the mobile terminal; and controlling the computing resources of the mobile terminal, the computing resources of the edge computing node, and the communication resources based on the resource scheduling decision.

[0099] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the resource scheduling method provided by the above methods. The method includes: obtaining a resource scheduling optimization model, wherein the resource scheduling optimization model is constructed with the optimization objective of minimizing the overall system cost constituted by the task queue backlog of the mobile terminal and the resource consumption of the mobile terminal and edge computing nodes, and with constraints including the availability of computing resources of the mobile terminal, the availability of computing resources of the edge computing nodes, and the availability of communication resources for task offloading of the mobile terminal; solving the resource scheduling optimization model to obtain resource scheduling decisions for the computing tasks of the mobile terminal; and controlling the computing resources of the mobile terminal, the computing resources of the edge computing nodes, and the communication resources based on the resource scheduling decisions.

[0100] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0101] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A resource scheduling method, characterized in that, include: A resource scheduling optimization model is obtained. The optimization objective of the resource scheduling optimization model is to minimize the overall system cost consisting of the task queue backlog of the mobile terminal and the resource consumption of the mobile terminal and the edge computing node. The model is constructed with the availability of computing resources of the mobile terminal, the availability of computing resources of the edge computing node, and the availability of communication resources for the mobile terminal to offload tasks as constraints. Solve the resource scheduling optimization model to obtain the resource scheduling decision for the computing tasks of the mobile terminal; Based on the resource scheduling decision, the computing resources of the mobile terminal, the computing resources of the edge computing node, and the communication resources are controlled.

2. The resource scheduling method according to claim 1, characterized in that, The availability of computing resources for the mobile terminal is determined based on the computing frequency of the mobile terminal; The availability of computing resources for the edge computing node is determined based on the computing frequency of the edge computing node; The availability of the communication resources is determined based on the proportion of the total communication bandwidth allocated to the mobile terminal.

3. The resource scheduling method according to claim 2, characterized in that, It also includes delay constraints; The latency constraint is determined based on the average length of the task queues of the mobile terminal and the edge computing node, respectively.

4. The resource scheduling method according to any one of claims 1 to 3, characterized in that, The steps for determining the overall system cost include: Based on the task queue length of the mobile terminal and the task queue length of the edge computing node, the system queue backlog is determined; Based on the computing and communication power consumption of the mobile terminal and the computing power consumption of the edge computing node, the system resource consumption is determined. The overall cost of the system is determined based on the system queue backlog and the system resource consumption.

5. The resource scheduling method according to claim 4, characterized in that, Solving the resource scheduling optimization model to obtain resource scheduling decisions for the computing tasks of the mobile terminal includes: Based on the resource scheduling optimization model, a single-slot optimization problem is constructed with the goal of minimizing the weighted sum of the changes in the system queue backlog and the system resource consumption. Obtain the system state information of the single-slot optimization problem; The system state information is input into a preset deep reinforcement learning network to obtain the resource scheduling decision output by the deep reinforcement learning network. The system status information includes current task queue information representing the backlog in the system queue, and channel status information representing the transmission quality of the communication resources.

6. The resource scheduling method according to any one of claims 1 to 3, characterized in that, The resource scheduling decision includes a task offloading ratio vector; The task offloading ratio vector is used to determine the proportion of data volume executed by the computing task in the mobile terminal, the airborne drone base station, and the ground macro base station.

7. The resource scheduling method according to any one of claims 1 to 3, characterized in that, The edge computing nodes include ground macro base stations deployed in terrestrial cellular networks and aerial drone base stations carried by drones.

8. A resource scheduling device, characterized in that, include: A construction unit is used to obtain a resource scheduling optimization model. The resource scheduling optimization model takes minimizing the overall system cost consisting of the task queue backlog of the mobile terminal and the resource consumption of the mobile terminal and the edge computing node as the optimization objective, and is constructed with the availability of computing resources of the mobile terminal, the availability of computing resources of the edge computing node, and the availability of communication resources for the mobile terminal to offload tasks as constraints. The solution unit is used to solve the resource scheduling optimization model to obtain the resource scheduling decision for the computing tasks of the mobile terminal; The control unit is used to control the computing resources of the mobile terminal, the computing resources of the edge computing node, and the communication resources based on the resource scheduling decision.

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 resource scheduling method as described in any one of claims 1 to 7.

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

Images